Академический Документы
Профессиональный Документы
Культура Документы
(Университет ИТМО)
На правах рукописи
Санкт-Петербург 2023
115
i=M
X
S̄ = arg min ||S(~ui )|| (3.4)
S2Eq k
i=1
8
XX <n , if tij = @nu
0 @ n1 x1 ... @ nk xd im , n 1
C(L u) = compl(tij ); compl(ti j) =
:0.5 , otherwise
j i
(3.5)
Введённые критерии C(L0 u) и Q(L0 u) для каждого из k уравнений систе�
мы образуют пространство для оптимизации. Так как приложение предполага�
ет использование эволюционного алгоритма многокритериальной оптимизации,
вводится отношение подчинения между кандидатными решениями оптимизаци�
онной задачи. Введём бинарное отношение на множестве созданных алгоритмом
систем, позволяющее определить предпочтительность одного решения над дру�
гим. Будем говорить, что кандидатное решение - система дифференциальных
уравнений S1 (u) доминирует по Парето над решением S2 (u) (обозначается как
S1 (u) S2 (u)), если для каждого i - индекса уравнения системы выполняется
Qi (S1 (u)) Qi (S2 (u)), и Ci (S1 (u)) Ci (S2 (u)), а также существует индекс j,
для которого Qj (S1 (u)) < Qj (S2 (u)) и/или Cj (S1 (u)) < Cj (S2 (u)). Подобное
отношение доминирования можно интерпретировать как факт того, что каж�
дое уравнение системы S1 (u) одновременно не хуже описывает динамическую
систему, и представлено при помощи более простой структуры.
Очевидно, что подобное отношение вводит лишь частичный порядок: для
S1 (u) и S2 (u), таких, что существует множество индексов критериев I1 , по ко�
торым система S1 является предпочительной, и I2 , по которым предпочтения
отдаются системе S2 , нельзя сказать, что одно кандидатное решение доминиру�
ет над другим. Множество кандидатных решений называется недоминируемым,
если для любых двух систем дифференциальных уравнений из этого множества
нельзя сказать, что одно находится в подчинённом отношении перед другим.
Кандидатное решение S0 (u) называется оптимальным (Парето-оптималь�
ным), если не существует иных решений S 0 (u), для которых выполняется
S 0 (u) S0 (u). Целью алгоритма является определение Парето-оптимального
недоминируемого множества кандидатных систем дифференциальных уравне�
ний: 8S 0 (u)9Si (u), Si (u) S 0 (u).
117
( 11 , 1
2, ... , ! ( 01
1
n_eq )
01 01
1 , 2 , ... , n_eq )
( 21 , 2
2, ... , ! ( 02
2
n_eq )
02 02
1 , 2 , ... , n_eq )
pi ⇠ U (0,1) (3.8)
if pi < pxover then 01 1
i = ↵ ⇤ i + (1 ↵) ⇤ 2i
else 01 1 02
i = i, i = i
2
dx
x sin t + cos t = 1 (4.1)
dt
b
Рисунок 4.2 � Предсказание состояния системы на основе полученных по
данным уравнений: левый график - корректное уравнение, правый график -
“переобученное” уравнение.
1 1 1
u = 10000 sin ( xy(1 x)(1 y))2 (4.8)
100 10 10
132
@u 1 1 1
= 1000 sin ( xy(1 x)(1 y))2 (4.9)
@t 100 10 10
@ 2u @ 2u @ 2u
= ↵ 1 + ↵ 2 (4.10)
@t2 @x2 @y 2
Алгоритм определения уравнения был настроен следующим образом: в ка�
честве механизма подготовки данных использовалась аппроксимации данных
при помощи полносвязной нейронной сети (4 скрытых слоя, содержащих 256,
64, 64, 1024 нейронов, и использующих гиперболический тангенс в качестве
функции активации, обучение на 10000 эпох). Последующее вычисление тен�
зоров производных исполнялось при помощи метода конечных разностей (цен�
тральная схема, шаг сетки - 0.01 ⇤ , где - шаг сетки, на которой были
поданы входные данные, по координатной оси дифференцирования). Поряд�
ки производных ограничивались 3-им по всем координатам. В эксперименте
был использован эволюционный алгоритм со следующими параметрами: число
эпох поиска уравнения nepochs = 25, размер популяции кандидатных решений
np op = 10, вероятность уравнения подвергнуться мутации - pmut = 0.2, до�
ля популяции, подвергаемая кроссоверу, nparent = 0.4, вероятности слагаемых
уравнения подвергнуться мутации внутри мутирующих кандидатных решений
и вероятность обмена между парой слагаемых в рамках кроссовера, составили
pterm_mut = pterm_crossover = 0.3. Для оценки приспособленности использовался
подход на основе
Результаты эксперимента таковы: метод успешно обнаруживает структу�
ру уравнения для интервала уровней шума до 7.5 %, что соответствует стан�
дартному отклонению гауссова шума в интервале [0, 0.2], умноженного на норму
поля во временном интервале. Ошибки весов в этом интервале незначительны,
как показано в Tab. 13. При более высоких уровнях шума (в интервале от 7,5%
до 10%) алгоритм обнаруживает дополнительные слагаемые, отсутствующие в
исходном уравнении, что приводит как к искажению структуры уравнения, так
и к некорректному вычислению весов. Наконец, при высоких уровнях (от 10%)
шума предлагаемый алгоритм теряет способность определять даже элементы
желаемой структуры уравнения, сходясь к структурам, описывающим шум в
данных.
133
@u @u @ 2u
+u =⌫ 2 =0 (4.11)
@t @x @x
@u @u @ 3 u
+ 6u + =0 (4.12)
@t @x @x3
p
c c
u= sech2 (x ct x0 ) (4.13)
2 2
Применение структуры к решению уравнения из соотношения 4.13, оце�
ниваемому на регулярной сетке, не позволило определить исходное уравнение
по данным. Неправильно обнаруженная модель возникает из-за более простых
случайных форм данных, таких как ut = cux , также обладающих низкими зна�
чениями ошибки уравнения. При использовании однокритериального подхода,
более низкая сложность этого уравнения приводит к большей вероятности его
открытия, чем для полного уравнения КдФ. Кроме того, отсутствие в структу�
ре производных высокого порядка, которые вычисляются с большей численной
ошибкой, чем производные первого и второго порядка, могут привести к более
низким значениям функционала ошибки, чем в корректном уравнении. Этот
эксперимент показывает, что однокритериальный алгоритм имеет склонность
134
1 @u @ 2 u @u
↵ +↵ 2 = (4.14)
r @r @r @t
2
Уровень шума 1 @u
r @r
@ u
@r2
@u
@t C
0 (1.5 ± ✏) · 10 7
(1.54 ± ✏) · 10 7
1 ✏
0.1 (1.51 ± ✏) · 10 7
(1.53 ± ✏) · 10 7
1 ✏
0.3 (1.4 ± 0.3) · 10 7
(1.5 ± 0.21) · 10 7
1 0.0023 ± 0.005
0.5 (1.45 ± 0.5) · 10 7
(1.5 ± 0.21) · 10 7
1 0.05 ± 0.026
0.7 (1.4 ± 0.7) · 10 7
(1.3 ± 0.4) · 10 7
1 0.1 ± 0.053
1 (1.3 ± 0.3) · 10 7
(1.1 ± 0.7) · 10 7
1 0.3 ± 0.1
Таблица 15 � Полученные коэффициенты перед слагаемыми уравнения
теплопроводности в цилиндрических координатах. Коэффициенты
нормализованы так, что перед слагаемым с @u @t коэффициент - единица. C -
соответствует свободному слагаемому.
2
8 1 @u 8@ u @u
(9.4 ± 0.11) · 10 + (9.423 ± 0.04) · 10 + ✏ ± 0.01 · 10 8
= (4.15)
r @r @r2 @t
1 @u @ 2u @u @u
4.1 · 10 8
· + 5.8 · 10 9
· 2 + v2 = (4.16)
r @r @r @r @t
136
8
>
> dx
= · (y x);
>
< dt
dy
dt = x · (⇢ z) y; (4.18)
>
>
>
: dz = xy
dt z;
104
101 102
100 10 1
2
6 × 100 10
4
10
0
4 × 10 10 6
0 8
3 × 10 10
10
10
2
2 × 10 0 10
Single Objective Multi-Objective Single Objective Multi-Objective Single Objective Multi-Objective
а б в
EPDE
SINDy
NL, % u0t uu0x u000
xxx
P, % b, µ ± 1.98 P b, µ ± 1.98 P b, µ ± 1.98 g.t. u0t + u000 0
xxx + 6uux = 0
0 100 1.001 ± 0.0 100 6.002 ± 0.0 100 1.06 ± 0.0 u0t + 0.992u000 0
xxx + 5.967uux = 0
0.5 80 0.913 ± 0.032 60 5.914 ± 2.59 70 1.31 ± 0.57 u0t 0.906u0x = 0
1 40 0.437 ± 0.156 0 0 u0t 0.816u0x = 0
2.5 100 0.36 ± 0.0 20 1.0 ± 0.0 20 0.01 ± 0.0 u0t 0.004u000
xxx 0.844u0x = 0
5 60 0.01 ± 2.13 · 10 5 80 1.0 ± 0.0 0 u0t 0.003u000
xxx 1.859uu0x + N [u] = 0
ного метода. Было показано, что при обучении метод позволяет получить кор�
ректное дифференциальное уравнение на входных данных с уровнями шума до
5-10 %, в то время как конкурирующие методы теряют возможность получить
уравнение на данных с уровнями шума около 1-2 %, т.к. точность получения
структуры составлет менее 50 %.
143
Заключение
ДУ - Дифференциальное уравнение
УрЧП - Дифференциальное уравнение в частных производных
ОДУ - Обыкновенное дифференциальное уравнение
ИНС - Искусственная нейронная сеть
LASSO - Least absolute shrinkage and selection operator, оператор наимень�
шего абсолютного сжатия и отбора
MAPE - Mean Absolute Percentage Error, cредняя абсолютная ошибка в
процентах
MOEA/DD - Evolutionary Many-Objective Optimization Algorithm Based
on Dominance and Decomposition
Уравнение КдВ - Уравнение Кортевега-де Фриза
147
Список литературы
20. Brunton S. L., Proctor J. L., Kutz J. N. Discovering governing equations from
data by sparse identification of nonlinear dynamical systems // Proceedings of
the National Academy of Sciences. � 2016.
21. Loiseau J. C., Brunton S. L. Constrained sparse Galerkin regression systems //
Journal of Fluid Mechanics. � 2018. � Т. 838. � С. 42�67.
22. Kaiser E., Kutz J. N., Brunton S. L. Sparse identification of nonlinear
dynamics for model predictive control in the low-data limit. � 2017. � URL:
https://arxiv.org/abs/1711.05501.
23. Sparse Identification of Nonlinear Dynamics for Rapid Model Recovery / M.
Quade [и др.]. � 2018. � URL: https://arxiv.org/abs/1803.00894v2.
24. Hirsh S. M., Barajas-Solano D. A., Kutz J. N. Sparsifying priors for Bayesian
uncertainty quantification in model discovery // Royal Society Open Science. �
2022. � Т. 9, № 2. � С. 211823.
25. Park J.-H., Dunson D. B. Bayesian generalized product partition model //
Statistica Sinica. � 2010. � С. 1203�1226.
26. Tran G., Ward R. Exact recovery of chaotic systems from highly corrupted
data // Multiscale Modeling and Simulation. � 2017. � Т. 15. � С. 1108�
1129.
27. Raissi M. Deep hidden physics models: Deep learning of nonlinear partial
differential equations. � 2018. � URL: https://arxiv.org/abs/1801.06637.
28. Berg J., Nystrom K. Data-driven discovery of PDEs in complex datasets. �
2018. � URL: https://arxiv.org/abs/1808.10788.
29. Berg J., Nystrom K. Neural network augmented inverse problems for PDEs. �
2017. � URL: https://arxiv.org/abs/1712.09685.
30. Pde-net: Learning pdes from data / Z. Long [и др.] // International Conference
on Machine Learning. � PMLR. 2018. � С. 3208�3216.
31. Long Z., Lu Y., Dong B. PDE-Net 2.0: Learning PDEs from data with a
numeric-symbolic hybrid deep network // Journal of Computational Physics. �
2019. � Т. 399. � С. 108925.
150
53. Alla A., Kutz J. N. Nonlinear model order reduction via dynamic mode
decomposition. � 2016. � URL: https://arxiv.org/abs/1602.05080.
54. Schmid P. J. Dynamic mode decomposition and its variants // Annual Review
of Fluid Mechanics. � 2022. � Т. 54. � С. 225�254.
55. Zhang Z. J., Duraisamy K. Machine learning methods for data-driven
turbulence modeling // 22nd AIAA Computational Fluid Dynamics
Conference. � 2015. � С. 2460.
56. Zhang Z., Singh A. New Approaches in Turbulence and Transition Modeling
Using Data-driven Techniques // AIAA Modeling and Simulation Technologies
Conference. � 2015.
57. Tracey B., Duraisamy K., Alonso J. Machine Learning Strategy to Assist
Turbulence Model Development // Proc. AIAA Scitech conference. � 2015.
58. Parish E., Duraisamy K. Quantification of Turbulence Modeling Uncertainties
Using Full Field Inversion // 15th AIAA Aviation Technology, Integration, and
Operations Conference. � 2015.
59. Hvatov A. Automated differential equation solver based on the parametric
approximation optimization // Mathematics. � 2023. � Т. 11, № 8. � С. 1787.
60. Ramm A., Smirnova A. On stable numerical differentiation // Mathematics
of computation. � 2001. � Т. 70, № 235. � С. 1131�1153.
61. Savitzky A., Golay M. J. Smoothing and differentiation of data by simplified
least squares procedures. // Analytical chemistry. � 1964. � Т. 36, № 8. �
С. 1627�1639.
62. Schmid M., Rath D., Diebold U. Why and how Savitzky–Golay filters should
be replaced // ACS Measurement Science Au. � 2022. � Т. 2, № 2. � С. 185�
196.
63. Johnson S. G. Notes on FFT-based differentiation // MIT Applied
Mathematics, Tech. Rep. � 2011.
64. Nix A. E., Vose M. D. Modeling genetic algorithms with Markov chains //
Annals of mathematics and artificial intelligence. � 1992. � Т. 5, № 1. �
С. 79�88.
153
Abstract
Data-driven methods provide model creation tools for systems where the appli-
cation of conventional analytical methods is restrained. The proposed method
involves the data-driven derivation of a partial di↵erential equation (PDE) for
process dynamics, helping process simulation and study. The paper describes
the methods that are used within the EPDE (Evolutionary Partial Di↵erential
Equations) partial di↵erential equation discovery framework [1]. The frame-
work involves a combination of evolutionary algorithms and sparse regression.
Such an approach is versatile compared to other commonly used data-driven
partial di↵erential derivation methods by making fewer assumptions about the
resulting equation. This paper highlights the algorithm features that allow data
processing with noise, which is similar to the algorithm’s real-world applications.
This paper is an extended version of the ICCS-2020 conference paper [2]
Keywords: data-driven modelling, PDE discovery, evolutionary algorithms,
sparse regression, spatial fields, physical measurement data
1. Introduction
2
162
60 2. Related work
3
163
4
164
equations.
95 Artificial neural networks provide a more versatile tool. This method is
based on the approximation of time derivative with combinations of spatial
derivatives and other functions. The ANN applications’ examples to the problem
of partial di↵erential equation discovery were presented in [8, 21, 22, 7, 6]. While
artificial neural networks can discover non-linear equations, they still rely on
100 approximating a determined term (time derivative of the first order), limiting
their flexibility.
3. Problem statement
The class of problems, which the described EPDE algorithm can solve, can
be summarized as follows: the process, which involves scalar field u, is occurring
105 in the area ⌦ and is governed by the partial di↵erential equation Eq. 1. How-
ever, there is no a priori information about the dynamics of the process except
that some form of PDE can describe it (for simplicity, we consider temporally
varying 2D field case, even though the problem could be formulated for an ar-
bitrary field). In recent developments, we have abandoned the assumption of
110 the constant weights in the partial di↵erential equations, allowing them to be
an arbitrary function (logarithmic, trigonometric) and thus expanding the class
of possible systems to study.
8
>
<F (u, @u @u @u @ 2 u @ 2 u @2u
@x1 , @x2 , ..., @t , @x21 , @x22 , ..., @t2 , ..., x) = 0;
(1)
>
:G(x) = 0, x 2 (⌦) ⇥ [0, T ];
5
165
120 defined limitations, including a range of the derivative orders, several terms in
the equations, and some factors in the term.
The resulting model Eq. 2 takes form of linear combination of terms, where
Q
Ntokens
each of them t(x) = j (with Ntokens is pre-defined algorithm hyper-
j=1
parameter) is constructed as the product of pre-computed elementary oper-
125 ators j, selected from a di↵erent groups of elementary operators of a same
nature. More detailed elementary operator i 2 = [j j; j - a group of
elementary operators of a same nature (for example, trigonometric group j =
@u
{sin(x1 ), cos(x1 ), sin(x2 ), ...}, or di↵erential operators group j = {u, @x , @u , ...}).
1 @x2
X
F (x) = ci ti (x) = 0; (2)
i
The addition of di↵erent groups j of terms allows to switch from the di↵er-
130 ential equation with the constant coefficients to di↵erential equations with the
variable coefficients.
The noise in this paper is assumed to be directional. The noise Exj in the
direction xj can be described as Eq. 3.
With x̄i “fixed” variables are denoted and ✏(xj ) is the noise, which in the
135 paper is assumed to be distributed normally N (0; ) with the expected value
of 0 and the variance of . It should be emphasized that poly-directional noise
forms as the superposition of the unidirectional noise operators, i.e. Exj ,xk =
Exj Exk . In what follows ¯ = max(u) is chosen as the multiplier of maximal
magnitude of the measured value in this direction. The noise level is defined in
140 the same way. In the text below bar over the variance, is omitted.
4. Method description
6
166
145 structure. The sparse regression aims to construct equation terms set, while the
evolutionary algorithm is focused on selecting significant terms from the created
set and calculating weights that will be present in the resulting equation. At
first, we introduce the preprocessing pipeline while later describe the algorithm
workflow.
To initialize the algorithm, time and spatial derivatives, which will later form
the desired equation, must be calculated. In specific situations, the derivatives
by themselves can be measured, and, therefore, this step can be skipped, but
often only the raw value of the studied function is available in the research. It
155 can be assumed without losing the generality that the measurements are held on
the rectangular (but not necessarily uniform) grid for the more straightforward
further computations. The multi-dimensional case requires more nuanced meth-
ods of obtaining derivatives, unlike the instances of a single dimension. In most
of the one-dimensional experiments, even on moderate noise levels, which can
160 be measured as Eq. 4, the finite-di↵erence method of derivative calculation can
lead to satisfactory results. It is important to note that the taken derivatives’
quality is crucial for acquiring the equation’s correct structure.
ku0 u ek2
Qnoise = ⇤ 100% (4)
ku0 k2
In general, the calculation of derivatives is the operation that is vulnerable
to the data’s noise. Also, the convergence of the algorithm and the resulting
165 equation depends on the input derivatives’ quality. If they are computed with
high errors, the resulting equation’s alterations can vary from incorrect coeffi-
cients to the entirely wrong structure. For these reasons, several noise-resistant
methods of partial derivative calculations have been introduced. Notably, they
include such commonly used methods as kernel smoothing [23], derivation of
170 polynomials, fitted over sets of points, and more uncommon ones like Kalman
filtering [24].
7
167
Z
ũ(s, t) = K (s s0 )u(s0 )ds0 ; (5)
2
1 1 X
K (s s0 ) = 2
exp ( 2
(s s0 )i ); (6)
2⇡ 2 i=1
Table 1: Noise levels (%) for the raw noised data and for the smoothed data
@u @2u
u @t @t2
8
168
190 the spatial derivative field, with the other spatial dimension processed by the
kernel.
Figure 1: Graph of a section over one spatial dimension for synthetic input function (solution
of wave equation with 2 spatial dimensions) in original state, with Gaussian noise, added to a
fraction (40%) of points of the domain, and after the noise was smoothed by Gaussian kernel
Di↵erentiation of the three di↵erent fields shown in Fig. 1 gives the derivative
fields that have values of the di↵erent orders. Thus, they are shown in the
di↵erent graphs in Fig. 2.
195 In most equations governing the real-world processes, derivatives’ orders are
limited to the first or second order. Derivatives of the slices shown in Fig. 1 are
represented in Fig. 2 and indicate that the proposed algorithm of noise reduction
in derivatives not only achieves values close to the values of the derivatives on
noiseless data but also preserves the structure of the fields, which is vital for the
200 main evolutionary algorithm due to the normalization of values on each time
frame.
9
169
Figure 2: Graph of first (left column) and second (right column) order time derivative, cal-
culated on input function (a)), noisy function (b)), and function with noise, smoothed by
Gaussian kernel (c)
10
170
function and its derivative along the x-axis is used as the regression feature set.
2 3 2 3 2 3
1 u (t0 , x0 ) ux (t0 , x0 )
6 . 7 6 .. 7 6 .. 7
6 . 7 6 7 6 7
6 . 7 6 . 7 6 . 7
6 7 6 7 6 7
6 7 6 7 6 7
f1 = 6 1 7 ; f2 = 6 u (ti , xj ) 7 ; f3 = 6 ux (ti , xj ) 7 ; ... (7)
6 7 6 7 6 7
6 . 7 6 .. 7 6 .. 7
6 .. 7 6 . 7 6 . 7
4 5 4 5 4 5
1 u (tm , xn ) ux (tm , xn )
2 3
u(t0 , x0 ) ⇤ ux (t0 , x0 )
6 .. 7
6 7
6 . 7
6 7
0 6 7
Fk = 6 u(ti , xj ) ⇤ ux (ti , xj ) 7 = f2 f3 ; (8)
6 7
6 .. 7
6 . 7
4 5
u(tm , xn ) ⇤ ux (tm , xn )
The normalization of terms values is held for each time frame passed into the
algorithm for the correct operation of regularized regression during the further
weight calculation phases. In this step, we can use arbitrary norm, but most
220 commonly, L2 norm or L1 norms are applied, as it is represented in Eq. 9.
11
171
inverse value of L2 -norm. The target term in the ”right side” of the equation
230 contains only one randomly selected term. Norm is taken as the di↵erences
between target Ftarget and the selected combination of features F with weighs
↵, obtained by the sparse regression (left side of the equation). Therefore, the
evolutionary algorithm’s task can be reduced to obtain the equation structure
with the highest fitness function value.
1
ff itness = ! max (10)
kF · ↵ Ftarget k2
235 The composition of the encoded terms represents the genotype of the indi-
vidual. These encodings contain the parameters of each token in the term. The
evolution of individuals is performed both by mutation and by a crossover in ev-
ery iteration step. The mutation for an individual is introduced as the random
change (addition, deletion, or alteration of factors) in its terms. For example,
240 this can result in shift of equation term ut to uxx ⇤ ut or ux ⇤ utt to ux ⇤ ut . The
elitism, introduced as the individual’s exclusion with the highest fitness value,
helps preserve the best-discovered candidate (the one with the highest fitness
function value) during the mutation step.
12
172
The main drawback of the LASSO regression is its disability to acquire the
correct values of the coefficients. Final linear regression over discovered e↵ective
terms is performed to obtain the resulting PDE’s actual coefficients. In the final
275 step, non-zero weights from the LASSO are rescaled with original unnormalized
data as features and the target.
The pseudo-code for the resulting algorithm is provided in Appendix A
5. Numerical experiments
280 a) Wave equation. The analysis of the algorithm performance is held on the
synthetic data. This simplification can show the result’s response to various
types and magnitudes of noise, which is generally unknown on the measurement
data. As in the previous studies, the solution of the wave equation with two
13
173
1 1 1
u = 10000 sin ( xy(1 x)(1 y))2 (12)
100 10 10
@u 1 1 1
= 1000 sin ( xy(1 x)(1 y))2 (13)
@t 100 10 10
The algorithm has proved to detect the correct structure of the equation
with the clean data, while on the noisy data, additional terms or completely
295 wrong structures have been detected.
14
174
weights calculation. Finally, with high noise levels, the proposed algorithm can
lose grasp of the equation’s correct structure.
Table 2: Discovered structures of the equations for the specific noise levels
In Fig. 3, the influence of the noise level added to the measured field on the
310 derivative fields is shown.
Figure 3: Noise levels of calculated first and second (dashed line) time derivatives, related to
noise levels of input data
In the other experiment with the same data set, the noise of relatively high
15
175
magnitudes was added to a minor fraction of points (5% of the total number). In
this case, the framework has shown similar results to the previous experiment:
until the noise level of approximately 15%, the discovered structure was correct.
315 On the data with higher noise magnitudes, errors in the structure of the equation
occurred. This experiment has shown that in the studied cases, the main limiting
factor for the algorithm’s performance with implemented preprocessing for noise
reduction is the noise level and not the distribution of noise across the studied
field.
320 b) Korteweg-de Vries equation’s solitary solution. To further analyze the syn-
thetic data’s algorithm performance, we have conducted additional experiments
on the Korteweg-de Vries equation and the heat transfer equation.
To create a more specific situation for the Korteweg-de Vries equation Eq. 15,
we have studied the solitary wave solution of the equation Eq. 16. This solution
325 represents the transfer of a single wave, propagation with speed c from the initial
position, specified by the wave crest’s location at x0 . The data for the test is
obtained from the solution function Eq. 16. The solution is evaluated on the
uniform grid of 101 spatial points in the interval x 2 [0, 10] and 151 time points
in the interval t 2 [0, 15].
@u @u @ 3 u
+ 6u + =0 (15)
@t @x @x3
p
c c
u= sech2 [ (x ct x0 )] (16)
2 2
330 The application of the framework to the solution Eq. 16, evaluated on a
regular grid, failed to rediscover the initial equation. An improperly discovered
model results from the simpler incidental forms in data, such as ut = cux . This
equation’s simplicity results in a higher probability of its discovery than for the
full KdV equation. Additionally, the absence of the high-order derivatives in
335 the structure, which are inevitably calculated with numerical error, may lead
to higher fitness function values than in the correct equation. This experiment
illustrates that the algorithm is susceptible to discovering “shortcut-equations”,
16
176
c) Heat equation with convection. To provide a more sophisticated test case for
the framework, we have utilized the convection-di↵usion equation. The equa-
tion belongs to the class of parabolic equations and has the structure Eq. 17,
345 where r represents gradient operator, and v is the velocity vector (field).
We have studied example of the equation with transfer in only one direction
(meaning,v = [v1 , 0]; v1 = 1, representing constant velocity field along x-
P @
axis), and ↵ = 1 - thermal di↵usivity of the medium; r = i ei @x i
- gradient
operator; ei - basis vector for i-th axis. The equation was solved on a grid with
350 100 ⇥ 100 ⇥ 100 nodes in domain between 0 and 10 along each axis. The ini-
tial and boundary conditions are correspondingly presented in the Eq. 12 and
Eq. 13.
@u
= r · (↵ru) v · (ru) (17)
@t
⇡ ⇡
u = 10 sin ( x(10 x)) sin ( x(10 x)) (18)
100 100
2⇡ ⇡
u = 10 sin ( t) sin ( y(10 y)) + 0.05t (19)
3 100
17
177
360 For the validation of the model, the dynamics of the two-dimensional field of
sea surface height (SSH) data from the NEMO ocean model for the Arctic region
(center of the Barents sea) for a modeling month a resolution of an hour has
been used. The area is known to have strong tides, leading to the discovery of
the time-dependent equation. It is necessary to emphasize that despite existing
365 Tidal equations, there is no single analytical equation for the specific case of
the SSH dynamics in this region due to the overlapping of di↵erent natures’
processes. The studied domain was divided into daily intervals to reduce the
risks of the deriving equation, which describes multiple processes following each
other in the domain. The data’s spatial properties are as follows: the intervals
370 between nodes are approximately 5 km, while the domain contained 50 ⇥ 50
nodes.
After the application of the framework to the data, we obtain the equation
in the form Eq. 20.
@u @u @2u
= 0.0506 0.0053 (20)
@x @t @t2
Eq. 20 was solved, and the calculated field was compared with the initial one
375 to validate the result of the algorithm. Since there is a second-order time deriva-
tive and first spatial derivative, the initial conditions (two initial time steps to
represent the field and its first time derivative for the beginning of the studied
period) and the boundary condition on one edge of the studied area are set. The
18
178
graphs of daily sea surface height dynamics from reanalysis and equation solu-
380 tion are presented in Fig. 4 and Fig. 5. The quality metrics show that the discov-
ered equation can describe the equation well: RM SE = 0.0434, M AE = 0.0446
for the field with values in interval between approximately 0.5 and 0.9.
Figure 4: Example of SSH field, obtained from reanalysis (upper row) and the same field from
Eq. 20, (lower row) for 3 time frames
@u @2u @2u @u @u
= ( 2 + 2 ) + u( + ) (22)
@t @x @y @x @y
19
179
Figure 5: Dynamics of sea surface height for September 18, 2013: reanalysis (denoted as data)
and solution of the equation, obtained from framework, denoted as model, for the center of
the studied area
390 Eq. 22 was solved using finite di↵erences, and the noise from normal dis-
tribution was added to simulate the previous experiment. The preprocessing
phase, described in previous sections and involving smoothing and derivatives
calculation, was performed to reduce noise’s influence on the resulting equation.
The added noise was created from k ⇥ maxx,y,t u(x, y, t) ⇥ N (0, 1). The
395 experiments resulting in the correct structure have been conducted with the
value k = 0.001, as in the compared study [7]. As the framework output, we
will consider the closest to the equation’s correct structure, obtained on the grid
of sparsity constant values.
These tests show that noise resistance corresponds with other framework
400 applications and is somewhat better than in the compared experiment. Despite
the insignificant di↵erence in the coefficient k (0.001 versus 0.00015), the noise
level di↵erence is significant (1% versus 6.3%). For the noise levels approxi-
mately below 5%, the correct equations were detected. The equation structure
deteriorates, which manifests in wrong weights and additional terms or lack of
405 mandatory ones.
20
180
Table 4: Discovered structures of the equations for the specific noise levels
21
181
inition of the tool mentioned above. For every token, the range for function
parameters (such as power or frequency) and markers are specified, setting the
behavior of functions in the equation structure (i.e., if a function can be present
in terms in the left side of the equation, if it is in the right part, or if multiple
425 tokens from a token family can be in a term).
The workflow of the main evolutionary algorithm, which forms the algo-
rithm’s cornerstone, is mutable via the evolutionary operator’s modifications.
To guide the operator’s development, we have introduced the builder class, rep-
resenting the eponymous pattern. The operators, presented in the Sec. 4, are
430 included in its default form, provided by the director class. However, the user
can modify all of the significant elements if the specific equation discovery task
requires it. The selection of the parameters for the evolutionary operators is
made in their definition in the builder class.
435 This section is dedicated to changing the di↵erentiation method. The pro-
posed algorithm has a modular structure. Thus, we may replace the di↵erentia-
tion algorithm from finite di↵erences or analytical di↵erentiation of polynomials
to the neural network approximation with further automatic di↵erentiation.
22
182
The wave equation’s test case with one spatial dimension was examined to
465 compare the noise-reducing properties of proposed derivatives evaluation meth-
ods: the kernel smoothing and analytical di↵erentiation of fitted polynomials
against the automatic di↵erentiation of artificial neural networks. The noise is
added in the way shown previously. However, data is not separated into sub-
domains. Here, we add the Gaussian noise with µ = 0 and = 0.03 ⇤ max(u(t))
470 to each of the time frames.
The lower quality of the artificial neural field reconstruction of the noisy
field presented in Fig. 6, can not be attributed to the overfitting or the under-
fitting of the network: the same structures are obtained in the initial stages
of loss function stabilization and on the latter epochs. Therefore, we can at-
475 tribute the method’s downside to the particular architecture: 3 fully connected
layers of the ANN with particular neurons. Nevertheless, the comparison of the
fields obtained after the di↵erentiation (see Fig. 7) shows that the default kernel
smoothing and the polynomial fitting algorithm can better calculate the fields.
23
183
Figure 6: The comparison between the solution of wave equation x and y axes are coordi-
nates, color value of the function u(x, y) a) and the ANN, fit to the solution b)
In the discussed experiments, the noise levels, introduced in Eq. 4, in the initial
480 fields are approximately 2.49%. The noise levels in the first time derivatives
(Fig. 8) are 31.6% and 1116.89% for default method and ANN accordingly, and
the first spatial derivatives (Fig. 9) are 29.7% and 927.14% correspondingly.
8. Conclusion
The proposed method has proven to be suitable for the data-driven deriva-
485 tion of equations that can model various physical processes. The robustness
of the algorithm to the noise in the input data provided by improved prepro-
cessing of data allows the framework applicable to real-world problems. Even
in the cases of substantial noise in the input data, the resulting equations had
the correct structures and, therefore, can correctly describe the studied system.
490 Other notable points about the algorithm operation can be stated:
24
184
(a) Field of the correct first time derivative (b) Field of the correct spatial time derivative
Figure 7: The fields of time and spatial derivatives, calculated from the noiseless data
(a) Kernel filtering & polynomial di↵erentiation (b) Automatic di↵erentiation result
Figure 8: The comparison between the di↵erentiation methods for first time derivative solution
of wave equation
25
185
(a) Kernel filtering & polynomial di↵erentiation (b) Automatic di↵erentiation result
Figure 9: The comparison between the di↵erentiation methods for first spatial derivative
solution of wave equation
26
186
Acknowledgements
References
[3] J. Berg, K. Nyström, Neural network augmented inverse problems for pdes,
arXiv preprint arXiv:1712.09685.
525 URL https://arxiv.org/abs/1712.09685
530 [5] S. H. Kang, W. Liao, Y. Liu, Ident: Identifying di↵erential equations with
numerical time evolution, arXiv preprint arXiv:1904.03538.
535 [7] Z. Long, Y. Lu, X. Ma, B. Dong, PDE-net: Learning PDEs from data, in:
International Conference on Machine Learning, 2018, pp. 3208–3216.
[8] M. Raissi, Deep hidden physics models: Deep learning of nonlinear partial
di↵erential equations, The Journal of Machine Learning Research 19 (1)
(2018) 932–955.
27
187
555 [14] M. Cabré, S. Solman, M. Nuñez, Creating regional climate change scenarios
over southern south america for the 2020’s and 2050’s using the pattern
scaling technique: validity and limitations., Climatic Change 98 (2010)
449–469. doi:10.1007/s10584-009-9737-5.
28
188
[23] I. Knowles, T. Le, A. Yan, On the recovery of multiple flow parameters from
transient head data, Journal of Computational and Applied Mathematics
169 (1) (2004) 1–15.
29
189
30
190
31
191
Abstract
While there are multiple approaches to simulating a dynamical system, that represents unknown
physical process, the majority of these methods can not be connected to the analytical equation-based
models. In this article, we suggest an approach of describing a process with a system of data-driven
ordinary di↵erential equations, that employs multiobjective optimization to obtain the set of candidate
systems. The main objective of the research is the development of a robust tool, that is able to construct
an analytical model for an arbitrary dynamical system, using as few assumptions as possible. To operate,
the algorithm demands data, describing state of the system, evolving in type. The optimization is held
in the criteria space of complexities and qualities of obtained di↵erential equations, allowing the selection
of the parsimonious model, that fits the needs of the process description. To improve the flexibility
of the algorithm, while sustaining the robustness to the noise in input data, linked to the real-world
applications, the algorithm performs simultaneous search of the systems’ equations, that have minimum
a priori assumptions about their structure. Thus, the model search algorithm obtains Pareto-optimal
set of systems. To validate the approach, experiments with the simple ”hunter-prey” model were held.
On this case the equation search algorithm has performed well, being able to correctly derive system of
equations with insignificant errors in coefficient estimations, linked to the numerical errors.
1 Introduction
Systems of ordinary or partial di↵erential equations (ODEs and PDEs) are powerful tools, that are able
to describe complex dynamics of structures, involving multiple variables. While there are many tools, that
can be used for creating mathematical models for processes, such as classical machine learning models, or
unconventional ones, like bayesian networks, they tend to have strict limitations to their applications. In cases
of many real-world systems, in addition to the aforementioned issues, these models are often abstracted from
the intrinsic physical principles, guiding the system. The classical approach to deriving systems of di↵erential
equations necessitates use of mathematical analysis in combination with the in-depth understanding of the
process. The data-driven approach to the system discovery involves creation of an individual di↵erential
equation for each dependent variable, that can be measured from a system.
The forms of the discovered models, that contain systems of di↵erential equations, are selected due to
the prevalence of di↵erential equation in physical systems. For example, flow of viscous fluid is governed by
Navier-Stokes equations, that are a system of partial di↵erential equations. Dynamics and interactions be-
tween electric and magnetic components of the electromagnetic field are described with Maxwell’s equations,
which are a system of PDE as well. Many simpler systems, such as rotation of the spherical pendulum, can
be defined with system of ordinary di↵erential equations.
Apart from the descriptive possibilities, provided by models in forms of systems of di↵erential equations,
obtained systems can be solved to predict further states of the process. While the toolkit for the automatic
solution of systems of ordinary/partial di↵erential equations is out of scope of this study, a number of stud-
ies have been conducted towards implementing equation-solving module into the frameworks of di↵erential
equations discovery as in work [1]. With this ability to solve model equations, the dynamics of the system
can be propagated into the future.
The problem of creating models for dynamical systems, governed by di↵erential equations, has seen surging
interest in the recent years. The first perspective to the task involves developing substitutes for the equations
in forms of propagation operators, that map the state of the system forward in time like in [2] or [3]. Dynamic
1
192
mode decomposition (DMD) involves approximation of the system dynamics with a finite-dimensional linear
operator. While that can be useful for multiple real-world applications, where the propagator is linear, many
other cases involve non-linear dynamics, that can not be fully explained with DMD approach.
A number of data-driven solutions to the problem of explaining dynamical system with explicitly derived
governing equations has been developed. Here, we will inspect methods, that are applicable not only for
problems of discovering ordinary di↵erential equations (ODE) and systems of ODEs, but also for tasks
of partial di↵erential equations discovery. The first problem has sufficient solution in forms of multilayer
stochastic models (MSMs), proposed in Kondrashov, Chekroun and Ghil in [4]. However, due to the non-
Markovian approach, the approach is not extendable to the problems of partial di↵erential equations.
The earliest advances were made with the symbolic regression [5]. Governing equations are viewed as
computational tree graphs, where leaves are inputs, and on the other levels various operators are located.
The search of the equation can be done with the typical graph-targeted evolutionary optimization algorithms.
More contemporary approaches are represented by sparse regression - based models, developed many works,
including Kaheman et. al. in [6] and Berg & Nystrom in [7], and with artificial neural networks (ANN)
representation of the dynamical system. While there are multiple approaches to discover di↵erential equations
with artificial neural networks, notable ones include PINN [8], PDE-Net, developed by Long et. al. [9], and
physics-informed neural networks by Raissi et. al. [10].
Partial di↵erential equation search with sparse regression uses LASSO operator, that is applied to ap-
proximate time derivative with a library of candidate terms. That library has to contain all of the possible
equation terms, and the usage of sparsity operator allows selection of only a few active feature terms. The
main issues of this approach can be linked with its rigidity: the term library has to be extensive enough to
contain all possible terms, including all of the non-linear functions, that can be present in equations. While
many of the presented approaches can be applied to the systems of di↵erential equations, their possibilities
are limited by description of time dynamics of a vector variable, like in paper by [11].
The algorithm, described in this article, is based on the multiobjective evolutionary optimization approach,
where the obtained model is evaluated by a number of metrics, describing quality and complexity of the
equations of the system. Thus, the algorithm is able to provide the parsimonious model, that is not overly
complex, but can sufficiently simulate the dynamics of the process. However, the problem of selecting
that parsimonious model from the discovered Pareto frontier is the problem for another study. This paper
is dedicated to the problem of discovering the optimal set of candidate equations, for the further expert
conclusions and applications.
The paper is divided into the following subsections: in the section Sec. 2 we formulate the equation
discovery problem; Sec. 3 provides generalized description of the developed algorithm; in Sec. 4 we evaluate
the performance of the algorithm on validational data. Sec. 5 concludes the paper and discusses further
development of the approach. All required data and code that are required to reproduce the results are
available from the URL placed in Sec. 6.
2
193
8 8 P Q
< L1 (u) = 0 < i ↵1i j t1ij = 0
S(u) = ... , ... (1)
: : P Q
Lk (u) = 0 i ↵ni j tnij = 0
The search for the optimal structures of equations in the system is done with the multi-objective opti-
mization, implemented with the Many-Objective Optimization Evolutionary Algorithm Based on Dominance
and Decomposition (MOEA/DD), introduced in [13].
The search is performed in the criteria space of complexities C(L0j u) and modelling errors Q(L0j u) for each
individual equations in a system. Therefore, the problem can be reformulated as Eq. 2. Here the constraints
are introduced in the equations construction logic rather than explicitly specified during the optimization
problem statement.
minimize F(S(u)) = (f1 (S(u)), ... , fm (S(u))) = (C(L01 u), Q(L01 u), ... , C(L0n u), Q(L0n u)) (2)
The complexity metric C(L0j u)is defined as a number of ”active” tokens in the equation, i.e. ones, that
are present in terms with non-zero coefficients.
The problem of selecting the most appropriate metric for evaluating the properties of process represen-
tation for the equation has been studied in work [1]. The best metrics for modelling quality are L2 norm
of matrices of di↵erential operator residuals, as presented in Eq. 3, or the norm of matrices of di↵erences
between the input variable fields and the solutions of corresponding equations Eq. 4.
3 Approach description
In this section we will briefly describe main diversions of our approach from the original algorithm, and case-
specific solutions employed during system of di↵erential equations derivation, such as evolutionary operators.
In accordance to the optimization objectives, stated in the previous section, algorithm performs simultaneous
search of system equations and parameters, which define the equations structures. The structure of an
equation can be decomposed into a set of equation terms and set of their real-valued coefficients ↵.
X Y
L0j u = ↵i tij (5)
i j
Q
The terms of constructed equations are represented with a product of tokens j tij , tij 2 T , elementary
building blocks, containing arbitrary user-defined functions. This approach enables discovery of non-linear
equations with compoundn structures, that have structures like in 5. During search of di↵erential equations
various derivatives (e.g. @@xuni , n 2 N) are included into the pool T . Other case-specific functions, or external
j
variables can be included as tokens into the token pool to be available for the algorithm during equation
search. For example, if the objective of a study is the discovery of the equation for the temperature dynamics
in a medium, than the velocity field of the medium can be considered as such external variable.
To create a system, that is able to fully model the studied process, it is possible make a an assumption,
that each equation of the system must represent spatial or temporal dynamics of at least one variable. By
the property of describing a dynamic of a variable we understand, that the equation contains corresponding
derivatives of the variable. During the evolutionary search, evolutionary operators, a↵ecting the structures
of the equations have to preserve the descriptive properties of such terms.
3
194
Figure 1: Scheme of the encoding for an ODE with sparsity constants as metaparameters.
To regulate the complexity of the equations, proposed by the algorithm, a regularization tool has to
be created. Its main objective is exclusion of terms with low significance and low descriptive power in the
resulting model. Selection of the terms can be done with sparse regression, operating with LASSO operator,
presented in eq. 6. As the predicted value of the operator, a random equation term, which represents an
”equation variable”, i.e. contains its derivative, is selected. LASSO operator is able to obtain a vector of
term weights with values of the terms in the left-hand side of the equation, evaluated on the space-time
grid, normalized and combined into matrix Fk , and vector of right-hand part values Ftarget,k . In the operator
statement, by || · ||i the i-th norm of matrix is meant.
The sparsity constant parameter , determines the penalty of optimized functional with respect to the
values of weights in , prioritizing setting zero coefficients to the less significant predictors. By regulating the
value of sparsity constant, the algorithm is able to control the complexity of the equation. Higher values of
promote equations with less numbers of terms, while lower values tend to lead to more complex equations.
Due to the significance of the sparsity parameter for the equation definition, it is included for each equation
in the system into the encoding of the individual.
The coefficients of the equation are computed with linear regression, where active terms from the left-hand
part are combined into matrix of predictors, and values of the term on the right part is used as a predicted
value.
4
195
4 Validation
To assess the performance of the proposed approach on tasks of discovering systems of equations, that govern
dynamical system, a number of validational experiments has been conducted. The most demonstrative
approach to check the behaviour our the algorithm employs synthetic data, obtained from the solution of
known equations.
A hunter-prey model, described by Lotka-Volterra equations (7), was selected as the dynamical system
to be described. The model represents simplified dynamics of two species: u = u(t) depicts ”prey”, while
v = v(t) represents the ”hunter” species. Constants ↵, , and determine the dynamics of the system. The
solutions of the equations were numerically obtained, using Runge-Kutta methods. The solutions for u(t)
and v(t) are demonstrated on figures Fig. 3. The process was modelled for a 1000 time steps with ini
⇢ 0
ut = ↵u uv
S(u) = (7)
vt0 = uv v
5
196
The Pareto-optimal set of equations, obtained from the algorithm, typically have forms, similar to the one,
presented in Fig. 4. Here the algorithm output is reformulated with the combination of optimized metrics:
instead of evaluating complexity or approximation errors of individual equations, they are viewed for the
system integrally. The allowable interval for complexity controlling parameters , used in sparsity operators,
is located between 10 8 and 10 2 . Next, additional family of trigonometric tokens was introduced into the
pool to create diversity of created terms.
While this test can not be considered as the comprehensive study of the algorithm properties, it can
be viewed as the proof of concept, that the algorithm is able to operate and discover the equations. 10
independent runs with 10 multiobjective optimization evolutionary algorithm iterations, and 10 more with
25 iterations were performed, and the obtained Pareto-optimal sets were compared. Due to the relatively
simple structure of the initial system of equations, a successful convergence to the similar (in terms of
obtaining sets with similar equations) was achieved in every case.
6
197
5 Conclusion
In this article, we proposed robust extension of the single di↵erential equation discovery approach to the
problems of creating models for systems of di↵erential equations. The multi-objective approach enables the
creation of a diverse set of models, and with the analysis of complexity-quality tradeo↵, an expert should
be able to select the parsimonious model for the process description. The main drawback of the developed
approach is high computational cost, which can be especially noticeable in cases of multidimensional data (i.e.
systems of partial di↵erential equations) or data with high noise levels, where high numbers of iterations are
required for the algorithm convergence. Therefore, improvement of the algorithm computational performance
can be stated as the priority for the further development. Also, the development of the sufficient tools for
using the derived equations for the process state prediction is another goal of the next researches. We hope,
that the research in this paper will motivate the research of the physics-based and interpretable methods of
machine learning and data-driven simulations.
Acknowledgements
This work was supported by the Analytical Center for the Government of the Russian Federation (IGK
000000D730321P5Q0002), agreement No. 70-2021-00141.
1 https://github.com/ITMO-NSS-team/EPDE
7
198
References
[1] M. Maslyaev and A. Hvatov, ”Solver-Based Fitness Function for the Data-Driven Evolutionary Discovery
of Partial Di↵erential Equations,” 2022 IEEE Congress on Evolutionary Computation (CEC), 2022.
[2] Brunton, S.L., Brunton, B.W., Proctor, J.L. et al. Chaos as an intermittently forced linear system. Nat
Commun 8, 19 (2017).
[3] P. J. Schmid and J. Sesterhenn. Dynamic mode decomposition of numerical and experimental data. In
61st Annual Meeting of the APS Division of Fluid Dynamics. American Physical Society, November
2008.
[4] D. Kondrashov, M. D. Chekroun and M. Ghil, “Data-driven non-Markovian closure models,” Physica
D: Nonlinear Phenomena, vol. 297, pp. 33-55, 2015.
[5] Schmidt Michael and Hod Lipson, “Distilling free-form natural laws from experimental data,” Science,
vol. 5923, pp. 81–85, 2019.
[6] K. Kaheman, J. N. Kutz and S. L. Brunton “SINDy-PI: a robust algorithm for parallel implicit sparse
identification of nonlinear dynamics,” Proceedings of the Royal Society A, vol. 476(2242), pp. 20200279,
2020.
[7] J. Berg and K. Nyström, “Data-driven discovery of PDEs in complex datasets,” Journal of Computa-
tional Physics, vol. 384, pp. 239-252, 2019.
[8] Han Gao, Matthew J. Zahr, Jian-Xun Wang, “Physics-informed graph neural Galerkin networks: A
unified framework for solving PDE-governed forward and inverse problems”, Computer Methods in
Applied Mechanics and Engineering, Volume 390, 2022, 114502, ISSN 0045-7825,
[9] Long, Z., et. al. (2018). “PDE-Net: Learning PDEs from Data”. Proceedings of the 35th International
Conference on Machine Learning, in Proceedings of Machine Learning Research, 80:3208-3216.
[10] M. Raissi, “Deep hidden physics models: Deep learning of nonlinear partial di↵erential equations,” The
Journal of Machine Learning Research, vol. 19(1), pp. 932-955, 2018.
[11] J. Zhang and W. Ma, “Data-driven discovery of governing equations for fluid dynamics based on molec-
ular simulation,” J. Fluid Mech, vol 892, pp. A5, 2020.
[12] F. Van Breugel, J. N. Kutz and B. W. Brunton, ”Numerical Di↵erentiation of Noisy Data: A Unifying
Multi-Objective Optimization Framework,” in IEEE Access, vol. 8, pp. 196865-196877, 2020.
[13] K. Li, K. Deb, Q. Zhang and S. Kwong, “An Evolutionary Many-Objective Optimization Algorithm
Based on Dominance and Decomposition,” in IEEE Transactions on Evolutionary Computation, vol. 19,
no. 5, pp. 694-716, Oct. 2015, doi: 10.1109/TEVC.2014.2373386.
[14] Mikhail Maslyaev, Alexander Hvatov, Anna V. Kalyuzhnaya, “Partial di↵erential equations discovery
with EPDE framework: Application for real and synthetic data”, Journal of Computational Science,
Volume 53, 2021, 101345, ISSN 1877-7503,
[15] I. Das and J. E. Dennis, “Normal-boundary intersection: A new method for generating the Pareto surface
in nonlinear multicriteria optimization problems,” SIAM J. Optim., vol. 8, no. 3, pp. 631–657, 1998.
8
199
Abstract—Partial differential equations provide accurate mod- of the dynamical system understanding, the researchers must
els for many physical processes, although their derivation can collect measurements as a preliminary part of the analysis.
be challenging, requiring a fundamental understanding of the Thus, the data-driven equation derivation can be introduced
modeled system. This challenge can be circumvented with the
data-driven algorithms that obtain the governing equation only as an alternative. Data-driven algorithms can develop a model
using observational data. One of the tools commonly used in for a process using measurements data and a few assumptions
search of the differential equation is the evolutionary optimization about the constructed model structure (i.e., the model will take
algorithm. In this paper, we seek to improve the existing the form of a partial differential equation).
evolutionary approach to data-driven partial differential equation One of the contemporary approaches to data-driven PDE
discovery by introducing a more reliable method of evaluating
the quality of proposed structures, based on the inclusion of the discovery is based on evolutionary algorithms (EA). Here, the
automated algorithm of partial differential equations solving. In objective of EA is the construction of a model in the form of a
terms of evolutionary algorithms, we want to check whether the partial differential equation from a selected set of elementary
more computationally challenging fitness function represented by constituents and with specific conditions that have the best
the equation solver gives the sufficient resulting solution quality performance on the input data. One of the issues linked with
increase with respect to the more simple one. The approach
includes a computationally expensive equation solver compared applying evolutionary operators to such ambiguously stated
with the baseline method, which utilized equation discrepancy to tasks is selecting an objective (fitness) function that shall be
define the fitness function for a candidate structure in terms of optimized during the problem solution.
algorithm convergence and required computational resources on The computation complexity of the fitness function is a
the synthetic data obtained from the solution of the Korteweg-de classical problem that is considered for classical optimization
Vries equation.
Index Terms—equation discovery, partial differential equation,
benchmark problems [1]. However, in the equation discovery
fitness function selection, data-driven modelling case, the function landscape cannot be known a priori. Thus,
only empirical conclusions on the influence of the hardness
I. I NTRODUCTION of computational complexity of a fitness function may be
obtained. In this paper, two possible approaches to compute
Differential equations are commonly used as mathematical fitness functions for equation discovery evolutionary algorithm
models for continuous physical processes. They describe the are considered. First is the computationally hard complete
interdependence of various derivatives of a studied multivariate equation solution. Equation solution should be a more noise-
function. In addition, differential equations can be used for resistant approach to evaluating the quality of evolutionary
system state predictions. By integrating the governing partial algorithm candidates due to the independence of a noisy
differential equation with specific initial and correctly stated data differentiation error. The ability to handle noisy data
boundary conditions, the state of the modeled process in the is crucial for modeling real-world physical systems since
future can be obtained. instrumental observational data gathering is usually imposed
The task of partial differential equation derivation for a by various disturbances. Moreover, the solver approach allows
specific problem in the past involved examining conservation considering boundary values, which are not in the scope of
laws that can be applied to the system and using analysis and previously used computationally more straightforward fitness
variational principles to extract the equation from the known calculation technique. It involved evaluating differential oper-
properties of the system. While these methods are widely used ator discrepancy from zero and performing poorly on noisy
to derive the equations, they demand significant preliminary data.
study of the process, which in some cases can not be held due The paper is structured as follows: Section II is dedicated to
to low comprehension of the system. Occasionally, regardless the analysis of existing approaches of data-driven discovery of
200
partial differential equations for purposes of modeling dynam- structures and systems of equations. Conceptually similar ap-
ical systems; Section III gives a brief formal overview of the proaches were introduced in [11] and [12]. However, they tend
solved problem and states the tasks for the research; Section IV to have stability and convergence issues while operating on
provides a description for the evolutionary algorithm of partial noisy data. The inaccurate estimations of derivatives, leading
differential equation discovery and compares the possible to the errors in calculating fitness functions and incorrect
ways of the case-specific fitness function definition; Section operation of selection operator, make their output inconsistent
V is dedicated to the analysis of algorithm performance on and the constructed models unreliable.
synthetic data, and Section VI outlines the paper and also During the equation discovery process, different forms of
describes the possible future work directions. operators appear. In a classical mathematical physics analysis,
many methods allow solving an equation of a given type
II. R ELATED WORK with a given type boundary conditions. Classical solution
The earliest advances in the data-driven algorithmic deriva- approaches require the expert to choose the proper method for
tion of partial differential equations were made with symbolic the given equation. Such an approach is improper for data-
regression [2]. The main idea of symbolic regression is con- driven discovery. We require a more automated yet maybe
structing an expression that describes dependencies in data less precise equation solver. Such approaches usually involve
using a computational graph. The leaves of the computational neural networks [13]. In the paper, we use an alternative
graph are used for modeled function and its derivatives, while realization of a neural network-based solver.
internal nodes are reserved for algebraic operations. Then, the
graph is optimized to represent the input data. While symbolic III. P ROBLEM STATEMENT
regression can construct an arbitrary form of expression for the This work’s primary goal is to check the influence of the
final model, it has some disadvantages, such as its tendency to more advanced fitness function on the data-driven derivation of
overtrain and vast search space, leading to poor performance ordinary and partial differential equations. As in the majority
of the optimization algorithm. Symbolic regression - based of the data-driven methods, the model is constructed based
approach has proved to be rather successful in problems of on sets of measurements. It has to be expected that a single
ordinary differential equation discovery. The examples of such dynamical process prevails in the entire area so that the derived
algorithm applications are numerous and include [3] and [4]. model can be applied to the complete dataset. The versatility
The former article involves application of numerical tools of of the approach makes possible the creation of models in forms
differential equations solution to evaluate the quality of a of non-linear ordinary and partial differential equations with
candidate, while the latter is notable for use of multiobjective the defined order n in the form of the expression Eq. 1, where
optimization, considering not only the solution fitting to the t is time, x is spatial coordinates vector, u is the function to
problem, but also its complexity. be modeled. The subscript denotes derivative along that axis,
The next class of data-driven methods of PDE discovery is e.g., ut means the first-order time derivative.
based on sparse regression, trained over a defined beforehand
library of candidate terms. Examples of works involving
sparse regression are numerous and include [5]–[8]. The main F (t, x1 , ... , xk , ut , ux1 , ... , utt , utx1 , ...) ⇠ Lu = 0 (1)
advantage of this approach is its low computational demands,
In contrast to the symbolic regression, which does not im-
but it can achieve such good performance with the cost of
pose any limitations on the models’ structures, the proposed al-
a limited expanse of possible equations, discoverable by the
gorithm makes assumptions about the equation form to reduce
framework: the researcher has to manually select all reasonable
the search space and avoid unlikely candidates. The candidate
terms for the equation.
model structure is granted a form of Eq. 2, where the factors
One of the most actively developing methods involves
fj belong to the set of derivatives {ut , ux1 , ... , utt , ...}.
applications of artificial neural networks (ANN) to derive the
The highest order of the derivative along the axis for the set
partial differential equation in the form of first-order time
may vary. In some cases, the dynamical system’s details may
derivative approximations. The most notable examples in this
indicate that the derivatives along a specific axis can be only up
include physics-inspired artificial neural networks, developed
to a certain order, or a similar notion can be obtained from an
by [9], aimed at the representation of both input data and the
algorithm’s launch results. Also, it is possible to assume that
structure of the PDE with ANNs. Another approach, proposed
the total number of significant (i.e., with nonzero coefficient)
by [10], views differential operators through the lens of con-
terms in the candidate equation and factors in a term will be
volution kernels, thus approximating the system’s dynamics.
relatively low due to the majority of existing equations for
The former approach has the downside of requiring knowledge
dynamical systems having this property.
about the general structure of the equation. The latter approach
can detect an arbitrary operator for the dynamics, but the P Q
L0 u = i ai (t, x)ci = 0 , ci = j fj ,
classes of learned equations are still limited by necessarily (2)
ai (t, x) = a0i (t, x) ⇤ bi , b 2 R
having first-time derivatives.
Evolutionary algorithms of PDE discovery proved to be While the process modeling shall be based on the solution
rather effective, being able to derive equations with complex of a single equation, the study [14] shows the benefits of
201
the multi-objective optimization approach. It allows the re- discovered equations are defined only by data and by the
searcher to obtain the set of Pareto-optimal solutions in terms hyperparameters of the algorithm, mainly sparsity parameters,
of multiple objective criteria, such as quality of the model which limit the number of terms in the equation.
(zero discrepancies or data reproduction precision computed Before the fitness evaluation discussion, we provide the gen-
with solver), its complexity, etc. For the latter criterion, the eral details of the evolutionary algorithm shall be introduced
approach adopts the proposed quantity of the significant terms without delving into the details of candidate equation quality
in the equation, containing derivatives or other variables, that evaluations. The search process for the best equation form is
are informative for the research, as shown in Eq. 3. The divided into several steps. At first, the evolutionary operators
selection of the former objective function is the main object are utilized to suggest a set of terms {a01 c1 , ...a0k ck }, where
of this research. ci is the product of selected derivatives or modeled function,
while the products a0i are constructed from an allowed set of
C(L0j u) = #(L0j u) (3) case-specific functions. Next, in a loop over the terms of the
equation in a set, the following procedure is performed: the
The multi-objective optimization operates by a separate EA term is selected as a ”right part of the equation”, while other
based on the MOEA/DD algorithm, proposed by [15]. In study ones are used to approximating it: the sparsity is applied, and
[14], the optimization problem was stated for the tasks of then the coefficients are refined. The overview of the EA is
differential equations systems, but its approach is applicable presented in the form of Alg. 1.
for the case of a single equation discovery.
In each algorithm launch, a total number of terms in the set
IV. A LGORITHM D ESCRIPTION {a01 c1 , ...a0k ck } is fixed for mutation and crossover evolution-
ary operators. Sparsity operator 4 has to be applied to set zero
This section describes the general details of the proposed coefficients to the additional terms in the set, absent in the
partial differential equations discovery approach. The gen- equation, describing the process. Here, the sparsity-promoting
eralized scheme of the developed evolutionary algorithm is technique, based on LASSO regression, is employed, although
presented in the Fig. 1. For the purposes of the fitness other works indicate that the correlation can be used as an
evaluation both optimization-based partial differential equation indicator of term significance. In the operator description, ⇤
solver and equation discrepancy evaluation procedure can be and refer to the sparse vector of real-valued intermediate
used. equation weights. The intermediate status is guided by the
nature of sparse regression, where the features and predicted
values have. is the sparsity constant, which is case of the
equation discovery regulates the number of terms with non-
zero coefficients. F is the matrix of left part terms, evaluated
on domain grid, while Ftarget is the vector of right part term.
Fig. 3. Scheme of the implemented evolutionary algorithm operators for tasks of partial differential equation discovery. As presented on the left, mutation
operator works by altering the existing solutions, while crossover operator, portrayed on the right, combines terms from selected equations.
Data: Encoded equation and boundary conditions, The data for the numerical solution was obtained using the
initial NN model model Wolfram Mathematica 12.3 software to avoid any numerical
Result: Trained neural network model that errors and shown in Fig. 4.
approximates the solution
Compute model Sobolev space norm min norm;
for NN in cache do
Train model to repeat NN output;
Apply differential operator to trained model;
Compute Sobolev space norm norm curr if
norm curr < min norm then
model=trained model ;
min norm = norm curr
else
pass
while patience < threshold do
Apply differential operator to trained model;
Compute Sobolev space norm norm ;
if norm oscilates near the same value then
patience = patience + 1
if norm is not improved in improving patience
steps then
patience = patience + 1
Gradient descent step for model with respect to Fig. 4. Visualization of the Korteweg-de Vries equation solution, used during
norm; the experiments
Algorithm 2: The pseudo-code of an equation solver
algorithm While, as it was stated earlier, the convergence on noiseless
data was proved to be rather robust, the additional Gaussian
noise (✏ ⇠ N (µ = 0; = n ⇤ ||u(t)||, n = 0.1, 0.2, ..., 0.9))
V. VALIDATION was added into the data for more detailed analysis of algorithm
The algorithm was tested on the synthetic data to evaluate performance. To assess the changes in algorithm behavior, the
the effects of the newly introduced PDE solver-based fitness success rate, manifesting in the fraction of algorithm launches,
function evaluation technique. The necessity of using synthetic when it converges to the correct equation (for one point on
data in the experiment is created by the fact, that for the the Pareto frontier), was utilized. The motivation behind the
validation purposes we need to know the true structure of selection of this metric was driven by the fact that it is
the equation. With this approach the performance of the consistent even for launches with different fitness function
algorithm on the experimental data can be evaluated with the evaluation approaches. The improvements were made only
success rate: fraction of independent launches, that result in in the selection operator, so no significant changes in the
the discovery of correct equations. By the correct results we coefficient calculation quality were made. For each of the noise
denote the equations with correct structure (i.e. set of terms), levels, 20 independent algorithm launches were commenced.
and with coefficients, that insignificantly (up to 5%) deviating The examples took forms of Pareto-frontiers, as presented in
form the ground truth. For these purposes, the solution of the Fig. 6
Korteweg-de Vries equation (Eq. 13) added was utilized. The From the success rate results, presented on the graph Fig. 5,
equation was solved on a grid of 31 ⇥ 31 points with the the interpretation can be made that on high-quality data, the
dimension t representing time and x - space. EA performance with both approaches tends to be similar.
205
entropy
Article
Towards Generative Design of Computationally Efficient
Mathematical Models with Evolutionary Learning
Anna V. Kalyuzhnaya *, Nikolay O. Nikitin, Alexander Hvatov, Mikhail Maslyaev, Mikhail Yachmenkov and
Alexander Boukhanovsky
Nature Systems Simulation Lab, National Center for Cognitive Research, ITMO University, 49 Kronverksky Pr.,
197101 St. Petersburg, Russia; nnikitin@itmo.ru (N.O.N.); alex_hvatov@itmo.ru (A.H.);
mikemaslyaev@itmo.ru (M.M.); mmiachmenkov@itmo.ru (M.Y.); boukhanovsky@mail.ifmo.ru (A.B.)
* Correspondence: anna.kalyuzhnaya@itmo.ru
Abstract: In this paper, we describe the concept of generative design approach applied to the
automated evolutionary learning of mathematical models in a computationally efficient way. To
formalize the problems of models’ design and co-design, the generalized formulation of the modeling
workflow is proposed. A parallelized evolutionary learning approach for the identification of
model structure is described for the equation-based model and composite machine learning models.
Moreover, the involvement of the performance models in the design process is analyzed. A set
of experiments with various models and computational resources is conducted to verify different
aspects of the proposed approach.
Keywords: generative design; automated learning; evolutionary learning; co-design; genetic programming
!"#!$%&'(!
!"#$%&'
way a complicated data-driven model that consists of many single models and feature
processing stages. Tasks in both examples can be formalized as the generative design of
computer models.
Both of cases (model as mathematical equation and complicated data-driven models)
have their own spheres of application, but they also can be joined as composite models. In
machine learning the composite model case often is described in terms of the multi-model
data-driven pipelines. If a single data-driven model cannot provide appropriate results,
various ensembling techniques like stacking or blending are applied [4]. To achieve better
quality, complex modeling pipelines can be used, that include different pre-processing
stages and can contain several types of models. A generalization of ensembling approaches
is the composite model concept [5]. A composite model has a heterogeneous structure, so
it can include models of different nature: machine learning (ML), equation-based, etc. [6].
A design of a composite model can be represented from an automated ML (AutoML)
perspective that may use a genetic algorithm for learning the structure. The evolutionary
learning approach seems to be a natural and justified choice because of several reasons.
First of all, the idea of generative design refers to the possibility of controlled open-ended
evolution under a set of restrictions. After that, genetic algorithms give flexible opportuni-
ties for treating mixed problems with combinatorial and real parts of a chromosome.
However, the design of the composite model may depend on different factors: the
desired modeling quality, computational constraints, time limits, interpreting ability re-
quirements, etc. It raises the problem of co-design [7] of the automatically generated
composite models with the specific environment. Generative co-design is an approach
which allows to synthesize jointly a set (mostly a pair) of objects that will be compatible
with each other. In context of this article these are mathematical models and computational
infrastructure. The conceptual difference between the generative design (that builds the
model on a basis of dataset only) and the generative co-design (that takes into account both
data and infrastructure) is illustrated in Figure 1. The structure of composite models can be
very complex, so it is complicated to construct the models in an expert way. For this reason,
different optimization techniques are used for the structural learning of the model. Usually,
the objective function for optimization is aimed to minimize the error of the predictions
obtained from the candidate model [8].
Model design
Graph Gene c
Input Quality measures
structure programming
data Output
Func onal Structural
complexity Machine learning data
blocks
measures
Computa onal
Infrastructure Quan ta ve Performance Hyperparameter
parameters measures op misa on
Genera ve co-design
Genera ve design
Figure 1. The description of the generative co-design concept: the different aspects of the model design (genotype,
phenotype, and the identification methods); the pipeline of the data-driven modeling; the difference between classical
design approach and co-design approach.
209
The paper is organized as follows. Section 2 describes the existing approaches to the
design of models. Section 3 provides the mathematical formulation for the model’s design
and co-design tasks and associated optimization problems. Section 4 described the actual
issues of generative co-design for the mathematical models. Section 5 provides the results
of experimental studies for different applications of generative design (composite models,
equation-based models, etc). The unsolved problems of co-design and potential future
works are discussed in Section 6. Section 7 provides the main conclusions.
2. Related Work
An extensive literature review shows many attempts for mathematical models design
in the different fields [9,10]. In particular, the methods of the automated model design
is highly valuable part of the various researches [11]. As an example, the equation-free
methods allow building the models that represent the multi-scale processes [12]. Another
example is building of the physical laws from data in form of function [13], ordinary
differential equations system [14], partial differential equations (PDE) [15]. The application
of the automated design of ML models or pipelines (which are algorithmicaly close notions)
are commonly named AutoML [8] although most of them work with models of fixed
structure, some give opportunity to automatically construct relatively simple the ML
structures. Convenient notation for such purpose is representation of a model as a directed
acyclic graph (DAG) [16]. Another example of popular AutoML tool for pipelines structure
optimization is TPOT [17].
To build the ML model effectively in the complicated high-performance environ-
ment [18], the properties of both algorithms and infrastructure should be taken into ac-
count. It especially important for the non-standard infrastructural setups: embedded [19],
distributed [20], heterogeneous [21] systems. Moreover, the adaptation of the model design
to the specific hardware is an actual problem for the deep learning models [22,23].
However, the application of co-design approaches [24] for the generative model
identification in the distributed or supercomputer environment [25,26] is still facing a
lot of issues. For example, the temporal characteristics of the designed models should
be known. The estimations of fitting and simulation time of the data-driven models can
be obtained in several ways. The first is the application of the analytical performance
models of the algorithm [27]. The identification of the analytical performance models can
be achieved using domain knowledge [28]. However, it can be impossible to build this
kind of model for the non-static heterogeneous environment. For this reason, the empirical
performance models (EPMs) are highly applicable to the different aspects of the generative
model design [29]. Moreover, the effective estimation of execution time is an important
problem for the generation of optimal computational schedule [30] or the mapping of
applications to the specific resources [31].
The execution of the complex resource-consuming algorithms in the specific infras-
tructure with limited resources raises the workflow scheduling problem [32]. It can be
solved using an evolutionary algorithm [33] or neural approaches [34].
It can be noted that the existing design and co-design approaches are mostly focused
on the specific application and do not consider the design for the different types of mathe-
matical models. In the paper, we propose the modified formulation of this problem that
allows applying the generative design and co-design approaches to the different tasks
and models.
3. Problem Statement
A problem of the generative design of mathematical models requires a model repre-
sentation as a flexible structure and appropriate optimization methods for maximizing a
measure of the quality of the designed model. To solve this optimization problem, different
approaches can be applied. The widely used approach is based on evolutionary algorithms
(e.g., genetic optimization implemented in TPOT [35] and DarwinML [16] frameworks) be-
cause it allows solving both exploration and exploitation tasks in a space of model structure
210
variants. The other optimization approaches like the random search of Bayesian optimiza-
tion also can be used, but the populational character of evolutionary methods makes it
possible to solve the generative problems in a multiobjective way and produce several
candidates model. Such formulation also can be successfully treated with the evolutionary
algorithms or hybrid ones that combine the use of evolutionary operators with additional
optimization procedures for increasing of robustness and acceleration of convergence. In
this section, we describe the problem of generative co-design of mathematical models and
computational resources in terms of the genetic programming approach.
A general statement for numerical simulation problem can be formulated as follows:
Y = H( M| Z ), (1)
S3 e3 e
A3 ... r
S1 S2 S3 ... S|s | a11 ... a1|A |,1 ... a1,|S | ... a|A |,|S | e1 ... er
p 1 p p p
Figure 2. The structure of the genotype during evolutionary optimization: functional properties, set of parameters and
relations between atomic blocks.
211
where f Q is a fitness function that characterizes quality generated mathematical and pQ max
is a maximal value of fitness function, model M is a space of possible model structures,
I + - actual computational resources, Tgen is time for model structure optimization critical
threshold tg . In such formulation we try to design the model with the highest quality, but
we need to rely optimization to single configuration of computational resources. This factor
is a strong limitation for the idea of generative design because this idea assumes flexibility
of searched solution including the possibility to find the most appropriate for applied task
combination of model structure and computational resources. The concept was illustrated
on Figure 1.
Model and computational infrastructure co-design may be formulated as follows:
with change of direct usage of infrastructure features to estimated time of model execution
via performance models TM ⇡ T = f T ( M, I ):
where f Q is single criteria fitness function that characterize goodness of fit of model with
additional limitations for expected model execution time TM and estimated time for
structural optimization Tgen .
Figure 3. Pareto frontier obtained after the evolutionary learning of the composite model in the “quality-execution time”
subspace. The points referred as M1 - M4 represent the different solutions obtained during optimization. pmax and tc
represent quality and time constraints.
In the context of automated models building and their co-design with computational
resources, performance models (PM) should be formulated as a prediction of expected
execution time with the explicit approximation
n o of a number of operations as a function
of computer model properties S, a1:|S| and infrastructure I parameters. However, for
different computer models classes, there are different properties of performance models. In
the frame of this paper, we address the following classes of models: ML models, numerical
models (based on the numerical solution of symbolic equations), and composite models
(that may include both ML and numerical models).
For ML models PM can be formulated as follows:
" #
PM OMLi,it
TML ( Z, M) = maxi  + O ( I ), (9)
it
Vi ( I ) + Consti ( I )
D n oE
According to structure M = S, E, a1:|S| for data driven-model case, duple hS, Ei
characterize structural
n o features of models (e.g., activation functions and layers in neural
networks) and a1:|S| characterize hyper-parameters.
For numerical models PM can be formulated as follows:
PM ONi
TNum ( R, M) = maxi + O ( I ), (10)
Vi ( I ) + Consti ( I )
where expected time prediction for each sequential part is based on properties of appropri-
ate model class: ⇢
OML, i f model is ML
OC = . (12)
ON, i f model is numerical
depend on the time-spatial grid taken, however, in another manner. It leads to the general
problem that the particular finite schemes cannot be used for the general equation solution.
The second approach is to approximate a solution with a neural network, which
somewhat mimics the finite element method. The neural networks are known as universal
approximators. However, their utility for differential equations solution is still arguable.
The main problem is that the good approximation of the field is not necessary leads to
the good derivative approximation [37]. There is a lot of workarounds to approximate
derivatives together with the initial field, however, it is done with the loss of generality.
The possible promising solution is to combine optimization methods, local neural
network approximation, and classical approach [38]. However, there is still a lot of the
“white spots”, since the arbitrary equation means a strongly non-linear equation with
arbitrary boundary conditions. Such a generality cannot be achieved at the current time
and requires a significant differentiation, approximation, and numerical evaluation method
development. The illustration examples of the inverse problem solution are shown in
Section 5.1.
can be formulated as search for generative evolutionary algorithm that is able to find
the best model structure M in limited time Tgen . This task can be solved by optimization
of hyper-parameters, evolutionary operators (and strategy of their usage) for generative
optimization algorithm and formulated as meta-optimization problem over a set of possible
algorithms U that are defined by a set of strategies B:
n o
U = {u( B)}, B = b1:| B| , b = h H, Ri, (13)
Figure 4. Setup that illustrates inefficiency of the parallel evolution implementation due to fitness function computation
complexity.
For the model’s generative design task, the most expensive step usually refers to the
evaluation of the fitness function value [6]. The calculation of the fitness function for the
individuals of the evolutionary algorithm can be parallelized in different ways that are
presented in Figure 5.
Individuals Individuals
[ X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, X11, X12 ] [ X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, X11, X12 ]
manager
Unit 0 Unit 3
Unit 0 Unit 3
Unit 1 worker
worker
worker
Figure 5. Approaches to the parallel calculation of fitness function with the evolutionary learning algorithm: (a) syn-
chronously, each element of the population is processed at one node until all is processed (b) asynchronously, one of the
nodes controls the calculations in other nodes.
216
The described approaches can be used for the different variants of the computational
environment used for the generation of the models. The practical application of the
generated models with the complex structure almost always difficult because of the high
computation complexity of the numerical model-based simulations.
There are several groups of models that can be separated by the simulation pipeline
structure. For the data-driven model, the computational cost of the fitting (identification)
stage is higher than for the simulation stage. For the equation-based numerical models
with rigid structure, there is no implicit fitting stage, but the simulation can be very expen-
sive. In practice, different parallelization strategies can be applied to improve simulation
performance [39].
The set of experiments that provides the examples to the problem raised in this issue
can be seen in Section 5.3.
T T
M
Data
Data
1 2 1 2 M
T Tuning step 2
T
1 1 3
T ing
ste
p T
Tun
The experiment that demonstrates the reduction of the search space for the composite
model design by the application of the modified hyperparameters tuning strategy for the
best individual is described in Section 5.4.
W
EPM
TNum ( Z|M, I ) = Â ww yw ( Z ), (15)
w =1
It is worth nothing that the obvious way to improve accuracy of estimation nON is to
use for experimental setup resources with characteristics of computational performance
close to V = meani (Vi ( I )) and task partitioning close to ON = meani (ONi ). Getting the
estimation of nON and infrastructure parameters Vi ( I ), Consti ( I ), O( I ) we may go to raw
estimation:
EPM ai nON ( M, Z )
TNum ( M, Z, I ) = maxi + O ( I ), (19)
Vi ( I ) + Consti ( I )
where ai is coefficient for model partitioning. Similar transformations could be made for
other models.
The experiments devoted to the identification of the empirical performance models
for both atomic and composite models are provided in Section 5.5.
218
5. Experimental Studies
The proposed approaches to the co-design of generative models cannot be recognized
as effective without experimental evaluation. To conduct the experiments, we constructed
the computational environment that includes and hybrid cluster and several multiprocessor
nodes that can be used to evaluate different benchmarks.
A set of experiments have been held with the algorithm of data-driven partial dif-
ferential equation discovery to analyze its performance with different task setups. All
experiments were conducted using the EPDE framework described in detail in [15].
The other set of experiments devoted to the automated design of the ML models
was conducted using the open-source Fedot framework (https://github.com/nccr-itmo/
FEDOT). The framework allows generating composite models using evolutionary ap-
proaches. The composite model generated by the framework can include different types
of models [6]. The following parameters of the genetic algorithm were used during the
experiments: maximum number of the generations in 20, number of the individuals in
each population is 32, probability of mutation, probability of mutation is 0.8, probability
of crossover is 0.8, maximum arity of the composite model is 4, maximum depth of the
composite model is 3. More detailed setup is described in [40].
256, 128, 64 neurons with sigmoid activation function. As the input data, the values of the
solution function for a wave equation (utt = au xx ), solved with the implicit finite-difference
method, have been utilized. Due to the nature of the implemented solution method, the
function values were obtained on the uniform grid. The training of ANN was done for a
specified number of epochs (500 for the conducted experiments), when of the each epoch
the training batch is randomly selected as a proportion of all points (0.8 of the total number
of points). To obtain the derivatives, the automatic differentiation methods, implemented
in the Tensorflow package are applied to the trained neural network.
0.8
0.6
0.4
y
0.2
approximation by ANN
0.0 solution of the equation
0 1 2 3 4 5
x
(a) (b)
Figure 7. Comparison of the equation solution and its approximation by artificial neural networks (ANNs) for a time slice
(a) and heatmap of the approximation error (u approx utrue ) (b).
Even with the presented very good approximation of the original field, the first
derivatives (Figure 8) are obtained with decent quality and may serve as the building
blocks. However, it is seen that the derivative field is significantly biased.
0.4
0.2
0.2
0.4
y
0.0
0.6
0.2
0.8
0.4
1.0
0 1 2 3 4 5 0 1 2 3 4 5
x x
(a) ut (b) u x
Figure 8. Comparison of derivatives obtained by polynomial differentiation and by symbolic regression for first time
derivative (a) first spatial derivatives (b) for a time slice (t = 50).
0.0
1.0
0.5
automatic differentiation by ANN
y
y
polynomial differentiation 0.5
1.0
1.5 0.0
2.0 0.5
2.5
1.0
0 1 2 3 4 5 0 1 2 3 4 5
x x
dx
L(t) = x (t) sin t + cos t = 1, (20)
dt
x (t) = sin t + C cos t. (21)
We have tried to rediscover the equation, based on data, obtained via analytical
differentiation of function (21), application of polynomial differentiation, and with the
derivative, calculated by automatic differentiation of fitted neural network. The series of
function values and the derivatives are presented in Figure 10. Here, we can see, that the
proposed ANN can decently approximate data; the analytical & polynomial differentiation
obtains similar fields, while automatic differentiation algorithm may result in insignificant
errors. 10 independent runs of the equation discovery algorithm have been performed
for each derivative calculation method, and the results with the lowest errors have been
compared. For the quality metric, the Mean Square Error of the vector, representing
the discrepancy of the function x̄ (t), which is the solution of discovered on data-driven
equation M(t) = 0 with aim of | M(t)| ! min, evaluated on the nodes of the grid was used.
While all of the runs resulted in the successful discovery of governing equations, the is-
sues with such equations are in the area of function parameters detection and calculating the
correct coefficients of the equation. The best result was achieved on the data from analytical
differentiation: MSE = 1.452 · 10 4 . The polynomial differentiation got the similar quality
MSE = 1.549 · 10 4 , while the automatic differentiation achieved MSE = 3.236 · 10 4 . It
could be concluded, that in the case of first-order equations, the error of the differentiation
has less order than all other errors and thus the fastest method for the given problem may
be used. However, in the PDE case, it is complicated to use only first-order derivatives,
whereas arbitrary ordinary differential equations may be represented as the system of the
first-order equations.
221
Figure 10. The solution of ODE from Equation (20), its approximation by neural network, and derivatives calculated by
analytic, polynomial and automatic differentiation.
∂2 u ∂2 u ∂2 u
2
= 2 + 2. (22)
∂t ∂x ∂y
222
However, that division has its downsides: smaller domains have less data, therefore,
the disturbances (noise) in individual point will have a higher impact on the results.
Furthermore, in realistic scenarios, the risks of deriving an equation, that describes a local
process, increases with the decrease in domain size. The Pareto front, indicating the trade-
off between the equation discrepancy and the time efficiency, could be utilized to find
the parsimonious setup of the experiment. On the noiseless data (we assume, that the
derivatives are calculated without the numerical error) even the data from a single point
will correctly represent the equation. Therefore, the experiments must be held on the data
with low, but significant noise levels.
We have conducted the experiments with the partition of data (Figure 11), containing
80 ⇥ 80 ⇥ 80 values, divided by spatial axes in fractions from the set {1, 10}. The experi-
ments were held with 10 independent runs on each of the setup (size of input data (number
of subdomains, into which the domain was divided, and sparsity constant, which affects
the number of terms of the equation).
100
60
40
20
0 20 40 60 80 100
Number of subdomains
(a) (b)
Figure 11. The results of the experiments on the divided domains. (a) evaluations of discovered equation quality for
different division fractions along each axis (2⇥ division represents division of domain into 4 square parts); (b) domain
processing time (relative to the processing of entire domain) for subdomain number.
The results of the test, presented in Figure 11, give insight into the consequences of
the processing domain by parts. It can be noticed, that with the split of data into smaller
portions, the qualities of the equations decrease due to the “overfitting” to the local noise.
However, in this case, due to higher numerical errors near the boundaries of the studied
domain, the base equation, derived from the full data, has its own errors. By dividing
the area into smaller subdomains, we allow some of the equations to be trained on data
with lower numerical errors and, therefore, have higher quality. The results, presented
in the Figure 11b are obtained only for the iterations of the evolutionary algorithm of the
equation discovery and do not represent the differences in time for other stages, such as
preprocessing, or further modeling of the process.
We can conclude that the technique of separating the domain into lesser parts and pro-
cessing them individually can be beneficial both for achieving speedup via parallelization
of the calculations and avoiding equations, derived from the high error zones. In this case,
such errors were primarily numerical, but in realistic applications, they can be attributed to
the faulty measurements or prevalence of a different process in a local area.
datasets were used as benchmarks that allow to analyze the efficiency of the generative
design in various situations.
To improve the performance of the model building (this issue was noted in Issue 2),
different approaches can be applied. First of all, caching techniques can be used. The cache
can be represented as a dictionary with the topological description of the model position in
the graph as a key and a fitted model as a value. Moreover, the fitted data preprocessor
can be saved in cache together with the model. The common structure of the cache is
represented in Figure 12.
Described by SID
Computa onally Shared storage for the ed models
expensive
(structural ID)
Cache dic onary
Iden ca on
Hyperparam. ( ng) SID 1 Cached model 1
Data-driven
model ... ...
Input data Predic on
SID N Cached model N
Depends on
Fast and simple
underlying chain Methods: append, clear, get
structure
Figure 12. The structure of the multi-chain shared cache for the fitted composite models.
The results of the experiments with a different implementation of cache are described
in Figure 13.
2000
1000
0
0 1 2 3 4 5 6 7 8 9 10
Generations
Figure 13. The total number model fit requests and the actually executed fits (cache misses) for the shared and local cache.
Local cache allows reducing the number of models fits up to five times against the
non-cached variant. The effectiveness of the shared cache implementation is twice as high
as that for the local cache.
The parallelization of the composite models building, fitting, and application also
makes it possible to decrease the time devoted to the design stage. It can be achieved
in different ways. First of all, the fitting and application of the atomic ML models can
be parallelized using the features of the underlying framework (e.g., Scikit-learn, Keras,
TensorFlow, etc [43]), since the atomic models can be very complex. However, this approach
is more effective in the shared memory systems and it is hard to scale it to the distributed
environments. Moreover, not all models can be efficiently parallelized in this way.
Then, the evolutionary algorithm that builds the composite model can be paralleled
itself, since the fitness function for each individual can be calculated independently. To con-
duct the experiment, the classification benchmark based at the credit scoring problem (https:
//github.com/nccr-itmo/FEDOT/blob/master/cases/credit_scoring_problem.py) was
224
used. The parameters of the evolutionary algorithm are the same as described at the
beginning of the section.
The obtained values of the fitness function for the classification problem are presented
in Figure 14.
(a) (b)
Figure 14. (a) The best achieved fitness value for the different computational configurations (represented as different
number of parallel threads) used to evaluate the evolutionary algorithm on classification benchmark. The boxplots are build
for the 10 independent runs. (b) Pareto frontier (blue) obtained for the classification benchmark in “execution time-model
quality” subspace. The red points represent dominated individuals.
(a) (b)
Figure 15. (a) The comparison of different scenarios of evolutionary optimization: best (ideal), realistic and worst cases
(b) The conceptual dependence of the parallelization efficiency from the variance of the execution time in population for
the different types of selection.
225
The same logic can be applied for the parallel fitting of the part of composite model
graphs. It raises the problem of the importance of assessment for the structural subgraphs
and the prediction of most promising candidate models before the final evaluation of the
fitness function will be done.
Figure 16. The comparison of different approaches to the evolutionary optimization of the composite models. The min-
max intervals are built for the 10 independent runs. The green line represents the static optimization algorithm with
20 individuals in the population; the blue line represented the dynamic optimization algorithm with 10 individuals in the
population. T0 , T1 and T2 are different real-time constraints, F0 , F1 and F2 are the values of fitness functions obtained with
the corresponding constraints.
226
Algorithm 1: The simplified pseudocode of the composite models tuning algorithm illustrated in Figure 6b.
Data: maxTuningTime, tuneData, paramsRanges
Result: tunedCompositeModel
fitData, validationData = Split(tuneData)
for atomicModel in compositeModel do
candidateCompositeModel = compositeModel
while tuningTime < maxTuningTime do
bestQuality = 0
candidateAtomicModel OptFunction(atomicModel, paramsRanges) // OptFunction can be
implemented as random search, Bayesian optimization, etc.
candidateCompositeModel Update(candidateCompositeModel, candidateAtomicModel)
Fit(candidateCompositeModel, fitData)
quality = EvaluateQuality (candidateCompositeModel, validationData)
if quality > bestQuality then
bestQuality = quality
bestAtomicModel = candidateAtomicModel
end
compositeModel Update(compositeModel, bestAtomicModel)
end
end
tunedCompositeModel = compositeModel
The results of the model-supported tuning of the composite models for the different
regression problems obtained from PMLB benchmark suite (Available in the https://
github.com/EpistasisLab/pmlb) are presented in Table 1. The self-developed toolbox
that was used to run the experiments with PMLB and FEDOT is available in the open
repository (https://github.com/ITMO-NSS-team/AutoML-benchmark). The applied
tuning algorithm is based on a random search in a pre-defined range.
Table 1. The quality measures for the composite models after and before random search-based tuning of hyperparameters. The
regression problems from PMLB suite [45] are used as benchmarks.
Benchmark Name MSE without Tuning MSE with Tuning R2 without Tuning R2 with Tuning
1203_BNG_pwLinear 8.213 0.102 0.592 0.935
197_cpu_act 5.928 7.457 0.98 0.975
215_2dplanes 1.007 0.001 0.947 1
228_elusage 126.755 0.862 0.524 0.996
294_satellite_image 0.464 0.591 0.905 0.953
4544_GeographicalOriginalofMusic 0.194 2.113 0.768 0.792
523_analcatdata_neavote 0.593 0.025 0.953 0.999
560_bodyfat 0.07 0.088 0.998 0.894
561_cpu 3412.46 0.083 0.937 0.91
564_fried 1.368 0.073 0.944 0.934
227
It can be seen that the hyperparameter optimization allow increasing the quality of
the models in most cases.
where T EPM —model fitting time estimation (represented in ms according to the scale of
coefficients from Table 2), Nobs —number of observations in the sample, N f eat —number of
features in the sample. The characteristics of the computational resources and hyperparam-
eters of the model are considered as static in this case.
We applied the least squared errors (LSE) algorithm to (23) and obtained the Q
coefficients for the set of models that presented Table 2. The coefficient of determination R2
is used to evaluate the quality of obtained performance models.
The application of the evolutionary optimization to the benchmark allows finding the
optimal structure of the composite model for the specific problem. We demonstrate EPM
constructing for the composite model which consists of logistic regression and random
forest as a primary nodes and logistic regression as a secondary node. On the basis of (11),
EPM for this composite model can be represented as follows:
2 N
Nobs
N f eat
EPM
TAdd = max (Q1,1 Nobs N f eat + Q2,1 Nobs , Q11,2 Nobs N f eat + Q2,2 Nobs ) + obs + , (24)
Q21,3 Q22,3
228
where TAddEPM —composite model fitting time estimated by the additive EMP, Q , j-i coeffi-
i
cient of j model type for EPM according to the Table 2.
The performance model for the composite model with three nodes (LR + RF = LR) is
shown in Figure 17. The visualizations for the atomic models are available in Appendix A.
Figure 17. Predictions of the performance model that uses an additive approach for local empirical performance models
(EPMs) of atomic models. The red points represent the real evaluations of the composite model as a part of validation.
where t is a variable of real time and rc is a critical threshold for values of error function
E. Such a problem is typical for models that are connected with a lifecycle of their pro-
totype, e.g., models inside digital shadow for industrial system [47], weather forecasting
models [48], etc.
Additional fitting of co-designed system may appear also on the level of model
execution where classic scheduling approach may be blended with model tuning. Classic
formulation of scheduling for resource intensive applications Tex min ( L⇤ ) = min G 0 ( L| M, I )
A
is based on idea of optimization search for such algorithm L⇤ that helps to provide minimal
computation time Tex min for model execution process through balanced schedules of
workload on computation nodes. However, such approach is restricted by assumption
of uniform performance models for all parts of application. In real cases performance of
application may change dynamically in time and among functional parts. Thus, to reach
more effective execution it is desirable to formulate optimization problem with respect to
possibility of tuning model characteristics that influence on model performance:
229
⇣n o⇤ ⌘ ⇣ ⇣n o⌘ ⌘ n o
Tex max a1:|S| , L⇤ = max G M a1:|S| , L| I , M = S⇤ , E⇤ , a1:|S| , L = { L m }, (26)
a,L
where G is objective function that characterize expected time of model execution with
respect to used scheduling algorithm L and model M. In the context of generative modeling
problem on the stage of execution model M can be fully described as a set of model
properties that consists of optimal model structure: optimal functions ⇤
n S o(from previous
stage) and additional set of performance influential parameters a1:|S| . Reminiscent
approaches can be seen in several publications, e.g., [49].
7. Conclusions
In this paper, we aimed to highlight the different aspects of the creation of mathe-
matical models using automated evolutionary learning approach. Such approach may be
represented from the perspective of generative design and co-design for mathematical
models. First of all, we formalize several actual and unsolved issues that exist in the
field of generative design of mathematical models. They are devoted to different aspects:
computational complexity, performance modeling, parallelization, interaction with the
infrastructure, etc. The set of experiments was conducted as proof-of-concept solutions
for every announced issue and obstacle. The composite ML models obtained by the FE-
DOT framework and differential equation-based models obtained by the EPDE framework
were used as case studies. Finally, the common concepts of the co-design implementation
were discussed.
Author Contributions: Conceptualization, A.V.K. and A.B.; Investigation, N.O.N., A.H., M.M. and
M.Y.; Methodology, A.V.K.; Project administration, A.B.; Software, N.O.N., A.H., and M.M.; Supervi-
sion, A.B.; Validation, M.M.; Visualization, M.Y.; Writing–original draft, A.V.K., N.O.N. and A.H. All
authors have read and agreed to the final publication of the manuscript.
Funding: This research is financially supported by the Ministry of Science and Higher Education,
Agreement #075-15-2020-808.
Conflicts of Interest: The authors declare no conflict of interest.
Abbreviations
The following abbreviations are used in this manuscript:
AI Artificial intelligence
ANN Artificial neural network
AutoML Automated machine learning
DAG Directed acyclic graph
EPM Empirical performance model
GPU Graphics processing unit
ML Machine learning
MSE Mean squared error
NAS Neural architecture search
ODE Ordinary differential equation
PDE Partial differential equation
PM Performance model
R2 Coefficient of determination
RMSE Root mean square error
ROC AUC Area under receiver operating characteristic curve
230
Table A1. Approximation errors for the different empirical performance models’ structures obtained
for the atomic ML models. The best suitable structure is highlighted with bold.
The visualization of the performance models predictions for the different cases is
presented in Figure A1. It confirms that the selected EPMs allow estimating the fitting time
quite reliably.
References
1. Packard, N.; Bedau, M.A.; Channon, A.; Ikegami, T.; Rasmussen, S.; Stanley, K.; Taylor, T. Open-Ended Evolution and Open-Endedness:
Editorial Introduction to the Open-Ended Evolution I Special Issue; MIT Press: Cambridge, MA, USA, 2019.
2. Krish, S. A practical generative design method. Comput.-Aided Des. 2011, 43, 88–100. [CrossRef]
3. Ferreira, C. Gene Expression Programming: Mathematical Modeling by an Artificial Intelligence; Springer: Berlin/Heidelberg, Germany,
2006; Volume 21.
4. Pavlyshenko, B. Using stacking approaches for machine learning models. In Proceedings of the 2018 IEEE Second International
Conference on Data Stream Mining & Processing (DSMP), Lviv, Ukraine, 21–25 August 2018; pp. 255–258.
5. Kovalchuk, S.V.; Metsker, O.G.; Funkner, A.A.; Kisliakovskii, I.O.; Nikitin, N.O.; Kalyuzhnaya, A.V.; Vaganov, D.A.; Bochenina,
K.O. A conceptual approach to complex model management with generalized modelling patterns and evolutionary identification.
Complexity 2018, 2018, 5870987. [CrossRef]
6. Kalyuzhnaya, A.V.; Nikitin, N.O.; Vychuzhanin, P.; Hvatov, A.; Boukhanovsky, A. Automatic evolutionary learning of composite
models with knowledge enrichment. In Proceedings of the 2020 Genetic and Evolutionary Computation Conference Companion,
Cancun, Mexico, 8–12 July 2020; pp. 43–44.
7. Lecomte, S.; Guillouard, S.; Moy, C.; Leray, P.; Soulard, P. A co-design methodology based on model driven architecture for real
time embedded systems. Math. Comput. Model. 2011, 53, 471–484. [CrossRef]
8. He, X.; Zhao, K.; Chu, X. AutoML: A Survey of the State-of-the-Art. arXiv 2019, arXiv:1908.00709.
9. Caldwell, J.; Ram, Y.M. Mathematical Modelling: Concepts and Case Studies; Springer Science & Business Media: Berlin/Heidelberg,
Germany, 2013; Volume 6.
10. Banwarth-Kuhn, M.; Sindi, S. How and why to build a mathematical model: A case study using prion aggregation. J. Biol. Chem.
2020, 295, 5022–5035. [CrossRef] [PubMed]
11. Castillo, O.; Melin, P. Automated mathematical modelling for financial time series prediction using fuzzy logic, dynamical
systems and fractal theory. In Proceedings of the IEEE/IAFE 1996 Conference on Computational Intelligence for Financial
Engineering (CIFEr), New York City, NY, USA, 24–26 March 1996; pp. 120–126.
12. Kevrekidis, I.G.; Gear, C.W.; Hyman, J.M.; Kevrekidid, P.G.; Runborg, O.; Theodoropoulos, C. Equation-free, coarse-grained
multiscale computation: Enabling mocroscopic simulators to perform system-level analysis. Commun. Math. Sci. 2003, 1, 715–762.
13. Schmidt, M.; Lipson, H. Distilling free-form natural laws from experimental data. Science 2009, 324, 81–85. [CrossRef]
14. Kondrashov, D.; Chekroun, M.D.; Ghil, M. Data-driven non-Markovian closure models. Phys. D Nonlinear Phenom. 2015,
297, 33–55. [CrossRef]
15. Maslyaev, M.; Hvatov, A.; Kalyuzhnaya, A. Data-Driven Partial Derivative Equations Discovery with Evolutionary Approach. In
International Conference on Computational Science; Springer: Berlin/Heidelberg, Germany, 2019; pp. 635–641.
16. Qi, F.; Xia, Z.; Tang, G.; Yang, H.; Song, Y.; Qian, G.; An, X.; Lin, C.; Shi, G. A Graph-based Evolutionary Algorithm for Automated
Machine Learning. Softw. Eng. Rev. 2020, 1, 10–37686.
17. Olson, R.S.; Bartley, N.; Urbanowicz, R.J.; Moore, J.H. Evaluation of a tree-based pipeline optimization tool for automating
data science. In Proceedings of the Genetic and Evolutionary Computation Conference, New York, NY, USA, 20–24 July 2016;
pp. 485–492.
18. Zhao, H. High Performance Machine Learning through Codesign and Rooflining. Ph.D. Thesis, UC Berkeley, Berkeley, CA,
USA, 2014.
19. Amid, A.; Kwon, K.; Gholami, A.; Wu, B.; Asanović, K.; Keutzer, K. Co-design of deep neural nets and neural net accelerators for
embedded vision applications. IBM J. Res. Dev. 2019, 63, 6:1–6:14. [CrossRef]
20. Li, Y.; Park, J.; Alian, M.; Yuan, Y.; Qu, Z.; Pan, P.; Wang, R.; Schwing, A.; Esmaeilzadeh, H.; Kim, N.S. A network-centric
hardware/algorithm co-design to accelerate distributed training of deep neural networks. In Proceedings of the 2018 51st Annual
IEEE/ACM International Symposium on Microarchitecture (MICRO), Fukuoka, Japan, 20–24 October 2018; pp. 175–188.
232
21. Bertels, K. Hardware/Software Co-Design for Heterogeneous Multi-Core Platforms; Springer: Berlin/Heidelberg, Germany, 2012.
22. Wang, K.; Liu, Z.; Lin, Y.; Lin, J.; Han, S. HAQ: Hardware-Aware Automated Quantization With Mixed Precision. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 16–20 June 2019.
23. Cai, H.; Zhu, L.; Han, S. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv 2018, arXiv:1812.00332.
24. Dosanjh, S.S.; Barrett, R.F.; Doerfler, D.; Hammond, S.D.; Hemmert, K.S.; Heroux, M.A.; Lin, P.T.; Pedretti, K.T.; Rodrigues, A.F.;
Trucano, T. Exascale design space exploration and co-design. Future Gener. Comput. Syst. 2014, 30, 46–58. [CrossRef]
25. Gramacy, R.B.; Lee, H.K. Adaptive Design of Supercomputer Experiments. 2018. Available online: http://citeseerx.ist.psu.edu/
viewdoc/download?doi=10.1.1.312.3750&rep=rep1&type=pdf (accessed on 26 December 2020).
26. Glinskiy, B.; Kulikov, I.; Snytnikov, A.V.; Chernykh, I.; Weins, D.V. A multilevel approach to algorithm and software design for
exaflops supercomputers. Numer. Methods Program. 2015, 16, 543–556.
27. Kaltenecker, C. Comparison of Analytical and Empirical Performance Models: A Case Study on Multigrid Systems. Master’s The-
sis, University of Passau, Passau, Germany, 2016.
28. Calotoiu, A. Automatic Empirical Performance Modeling of Parallel Programs. Ph.D. Thesis, Technische Universität, Berlin,
Germany, 2018.
29. Eggensperger, K.; Lindauer, M.; Hoos, H.H.; Hutter, F.; Leyton-Brown, K. Efficient benchmarking of algorithm configurators via
model-based surrogates. Mach. Learn. 2018, 107, 15–41. [CrossRef]
30. Chirkin, A.M.; Belloum, A.S.; Kovalchuk, S.V.; Makkes, M.X.; Melnik, M.A.; Visheratin, A.A.; Nasonov, D.A. Execution time
estimation for workflow scheduling. Future Gener. Comput. Syst. 2017, 75, 376–387. [CrossRef]
31. Gamatié, A.; An, X.; Zhang, Y.; Kang, A.; Sassatelli, G. Empirical model-based performance prediction for application mapping
on multicore architectures. J. Syst. Archit. 2019, 98, 1–16. [CrossRef]
32. Shi, Z.; Dongarra, J.J. Scheduling workflow applications on processors with different capabilities. Future Gener. Comput. Syst.
2006, 22, 665–675. [CrossRef]
33. Visheratin, A.A.; Melnik, M.; Nasonov, D.; Butakov, N.; Boukhanovsky, A.V. Hybrid scheduling algorithm in early warning
systems. Future Gener. Comput. Syst. 2018, 79, 630–642. [CrossRef]
34. Melnik, M.; Nasonov, D. Workflow scheduling using Neural Networks and Reinforcement Learning. Procedia Comput. Sci. 2019,
156, 29–36. [CrossRef]
35. Olson, R.S.; Moore, J.H. TPOT: A tree-based pipeline optimization tool for automating machine learning. Proc. Mach. Learn. Res.
2016, 64, 66–74.
36. Evans, L.; Society, A.M. Partial Differential Equations; Graduate Studies in Mathematics; American Mathematical Society:
Providence, RI, USA, 1998.
37. Czarnecki, W.M.; Osindero, S.; Jaderberg, M.; Swirszcz, G.; Pascanu, R. Sobolev training for neural networks. In Proceedings of
the Advances in Neural Information Processing Systems 30 (NIPS 2017), Long Beach, CA, USA, 4–9 December 2017; pp. 4278–4287.
38. Raissi, M.; Perdikaris, P.; Karniadakis, G.E. Physics-informed neural networks: A deep learning framework for solving forward
and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 2019, 378, 686–707. [CrossRef]
39. Epicoco, I.; Mocavero, S.; Porter, A.R.; Pickles, S.M.; Ashworth, M.; Aloisio, G. Hybridisation strategies and data structures for
the NEMO ocean model. Int. J. High Perform. Comput. Appl. 2018, 32, 864–881. [CrossRef]
40. Nikitin, N.O.; Polonskaia, I.S.; Vychuzhanin, P.; Barabanova, I.V.; Kalyuzhnaya, A.V. Structural Evolutionary Learning for
Composite Classification Models. Procedia Comput. Sci. 2020, 178, 414–423. [CrossRef]
41. Full Script That Allows Reproducing the Results Is Available in the GitHub Repository. Available online: https://github.
com/ITMO-NSS-team/FEDOT.Algs/blob/master/estar/examples/ann_approximation_experiments.ipynb (accessed on
26 December 2020).
42. Full Script That Allows Reproducing the Results Is Available in the GitHub Repository. Available online: https://github.com/
ITMO-NSS-team/FEDOT.Algs/blob/master/estar/examples/Pareto_division.py (accessed on 26 December 2020).
43. Géron, A. Hands-on Machine Learning with Scikit-Learn, Keras, and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent
Systems; O’Reilly Media: Sebastopol, CA, USA, 2019.
44. Nikitin, N.O.; Vychuzhanin, P.; Hvatov, A.; Deeva, I.; Kalyuzhnaya, A.V.; Kovalchuk, S.V. Deadline-driven approach for multi-
fidelity surrogate-assisted environmental model calibration: SWAN wind wave model case study. In Proceedings of the Genetic
and Evolutionary Computation Conference Companion, Prague, Czech Republic, 13–17 July 2019; pp. 1583–1591.
45. Olson, R.S.; La Cava, W.; Orzechowski, P.; Urbanowicz, R.J.; Moore, J.H. PMLB: A large benchmark suite for machine learning
evaluation and comparison. BioData Min. 2017, 10, 1–13. [CrossRef]
46. Li, K.; Xiang, Z.; Tan, K.C. Which surrogate works for empirical performance modelling? A case study with differential evolution.
In Proceedings of the 2019 IEEE Congress on Evolutionary Computation (CEC), Wellington, New Zealand, 10–13 June 2019;
pp. 1988–1995.
47. Bauernhansl, T.; Hartleif, S.; Felix, T. The Digital Shadow of production–A concept for the effective and efficient information
supply in dynamic industrial environments. Procedia CIRP 2018, 72, 69–74. [CrossRef]
48. Cha, D.H.; Wang, Y. A dynamical initialization scheme for real-time forecasts of tropical cyclones using the WRF model. Mon.
Weather Rev. 2013, 141, 964–986. [CrossRef]
49. Melnik, M.; Nasonov, D.A.; Liniov, A. Intellectual Execution Scheme of Iterative Computational Models based on Symbiotic
Interaction with Application for Urban Mobility Modelling. IJCCI 2019, 1, 245–251.
233
ABSTRACT some information about its operation. In the case of modeling phys-
Evolutionary di�erential equation discovery proved to be a tool to ical processes, commonly, the most suitable models have forms of
obtain equations with less a priori assumptions than conventional partial di�erential equations. Thus many recent studies aimed to
approaches, such as sparse symbolic regression over the complete develop the concept of data-driven di�erential equations discovery.
possible terms library. The equation discovery �eld contains two In the paper, data-driven discovery implies obtaining a di�eren-
independent directions. The �rst one is purely mathematical and tial equation from a set of empirical measurements, describing the
concerns di�erentiation, the object of optimization and its relation dynamics of a dependent variable in some domain. Furthermore,
to the functional spaces and others. The second one is dedicated equation-based models can be incorporated into pipelines of au-
purely to the optimizatioal problem statement. Both topics are tomated machine learning, that can include arbitrary submodels,
worth investigating to improve the algorithm’s ability to handle with approach, discussed in paper [14].
experimental data a more arti�cial intelligence way, without signif- Initial advances in di�erential equations discovery were made
icant pre-processing and a priori knowledge of their nature. In the with symbolic regression algorithm, as in [1]. The algorithm em-
paper, we consider the prevalence of either single-objective opti- ploys genetic programming to detect the graph, that represents
mization, which considers only the discrepancy between selected di�erential equation. One of the groups of the most simple yet
terms in the equation, or multi-objective optimization, which addi- practical techniques of equation construction is based on the sparse
tionally takes into account the complexity of the obtained equation. linear regression (least absolute shrinkage and selection operator),
The proposed comparison approach is shown on classical model introduced in works [11], [15], [16], and other similar projects. This
examples – Burgers equation, wave equation, and Korteweg - de approach has limited �exibility, having applicability restrictions
Vries equation. in cases of the equation with low magnitude coe�cients, being
discovered on noisy data. This issue is addressed by employing
CCS CONCEPTS Bayesian interference as in [12] to estimate the coe�cients of the
equation, as in work [4]. To account for the uncertainty in the
• Applied computing ! Mathematics and statistics; • Computing
resulting model, the approximating term library can be biased sta-
methodologies ! Heuristic function construction.
tistically [2]. Physics-informed neural networks (PINN) form the
next class of data-driven equation discovery tools, representing
KEYWORDS
the process dynamics with arti�cial neural networks. The primary
symbolic regression, dynamic system modeling, interpretable learn- research on this topic is done in work [13], while recent advances
ing, di�erential equations, sparse regression have been made in incorporating more complex types of neural
ACM Reference Format: networks in the PINNs [3, 17].
Mikhail Maslyaev and Alexander Hvatov. 2023. Comparison of Single- and In recent studies [7, 10], evolutionary algorithms have proved
Multi- Objective Optimization Quality for Evolutionary Equation Discovery. to be a rather �exible tool for di�erential equation discovery, de-
In Genetic and Evolutionary Computation Conference Companion (GECCO manding only a few assumptions about the process properties. The
’23 Companion), July 15–19, 2023, Lisbon, Portugal. ACM, New York, NY, problem is stated as the process representation error minimization.
USA, 4 pages. https://doi.org/10.1145/3583133.3590601 Implementing multi-objective evolutionary optimization, �rst in-
troduced for DE systems, as in [8], seems to be a feasible way to
1 INTRODUCTION improve the quality of the equation search, operating on fewer
The recent development of arti�cial intelligence has given high initial assumptions and providing higher diversity among the pro-
importance to problems of interpretable machine learning. In many cessed candidates. Additional criteria can represent other valuable
cases, users value models not only for their quality of predicting properties of the constructed models, namely conciseness.
the state of the studied system but also for the ability to provide This study compares the performance of single- and multi- objec-
tive optimization. Namely, the hypothesis that the multi-objective
Permission to make digital or hard copies of part or all of this work for personal or optimization creates and preserves diversity in the population and
classroom use is granted without fee provided that copies are not made or distributed
for pro�t or commercial advantage and that copies bear this notice and the full citation thus may achieve a better �tness function values, than that of a
on the �rst page. Copyrights for third-party components of this work must be honored. single-objective approach.The theoretical comparison shows that
For all other uses, contact the owner/author(s). multi-objective algorithms allow escaping local minima as soon as
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal
© 2023 Copyright held by the owner/author(s). the number of objectives is reasonably small [5]. For equation dis-
ACM ISBN 979-8-4007-0120-7/23/07. covery applications, the function landscapes have a more complex
https://doi.org/10.1145/3583133.3590601
234
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal M. Maslyaev, and A. Hvatov
structure, so increased diversity of the population can bene�t the 2.2 Mechanics of implemented evolutionary
resulting quality. operators
To direct the search for the optimal equations, standard evolution-
2 ALGORITHM DESCRIPTION ary operators of mutation and cross-over have been implemented.
The data-driven di�erential equation identi�cation operates on While the mechanics of single- and multi-objective optimization
problems of selecting a model for dynamics of the variable D = in the algorithm di�er, they work similarly on the stage of apply-
D (C, x) in a spatio-temporal domain (0,) ) ⌦, that is implicitly
>
ing equation structure-changing operators. With the graph-like
described by di�erential equation Eq. 1 with corresponding initial encoding of candidate equations, the operators can be represented
and boundary conditions. It can be assumed, that the order of the as changes, introduced into its subgraphs.
unknown equation can be arbitrary, but rather low (usually of The algorithm properties to explore structures are provided by
second or third order). mutation operators, which operate by random token and term ex-
mD mD mD changes. The number of terms to change has no strict limits. For
(C, x, D,
, , ... )=0 (1) tokens with parameters (?:+1, ... ?= ) 2 R= : , such as a para-
mC mG 1 mG=
Both multi-objective and single-objective approaches have the metric representation of an unknown external dependent variable,
same core of "graph-like" representation of a di�erential equation parameters are also optimized: the mutation is done with a random
(encoding) and similar evolutionary operators that will be described Gaussian increment.
further. In order to combine structural elements of better equations,
the cross-over operator is implemented. The interactions between
2.1 Di�erential equation representation parent equations are held on a term-level basis. The sets of terms
pairs from the parent equation are divided into three groups: terms
To represent the candidate di�erential equation the computational
identical in both equations, terms that are present in both equations
graph structure is employed. A �xed three-layer graph structure is
but have di�erent parameters or only a few tokens inside of them
employed to avoid the infeasible structures, linked to unconstrained
are di�erent, and the unique ones. The cross-over occurs for the two
graph construction and overtraining issues, present in symbolic
latter groups. For the second group it manifests as the parameter
regression. The lowest level nodes contain tokens, middle nodes
exchange between parents: the new parameters are selected from
and the root are multiplication and summation operations. The
the interval between the parents’ values.
data-driven equations take the form of a linear combination of
Cross-over between unique terms works as the complete ex-
product terms, represented by the multiplication of derivatives,
change between them. The construction of exchange pairs between
other functions and a real-valued coe�cient Eq. 2.
these tokens works entirely randomly.
(
0 (C, x, D, mD , mD , ... mD ) = Õ U Œ 5 = 0
mC mG 1 mG= 8 8 9 89
(2) 2.3 Optimization of equation quality metric
⌧ 0 (D)| = 0
The selection of the optimized functional distinguishes multiple
Here, the factors 58 9 are selected from the user-de�ned set of approaches to the di�erential equation search. First of all, a more
elementary functions, named tokens. The problem of an equation trivial optimization problem can be stated as in Eq. 4, where we
search transforms into the task of detecting an optimal set of tokens assume the identity of the equation operator 0 (D) = 0 to zero as
to represent the dynamics of the variable D (C, x), and forming the in Eq. 2.
equation by evaluating the coe�cients U = (U 1, ... U< ).
During the equation search, we operate with tensors of token ’ ÷
values, evaluated on grids DW = D (CW , xW ) in the processed domain &>? ( 0 (D)) = || 0 (D)||= = || U8 58 9 ||= ! min (4)
> U 8 C8 9
(0,) ) ⌦. 8 9
Sparsity promotion in the equation operates by �ltering out
An example of a more complex optimized functional is the norm
nominal terms with low predicting power and is implemented with
of a discrepancy between the input values of the modelled variable
LASSO regression. For each individual, a term (without loss of
and the solution proposed by the algorithm di�erential equation,
generality, we can assume that it is the <-th term) is marked to be a
estimated on the same grid. Classical solution techniques can not
"right-hand side of the equation" for the purposes of term �ltering
Œ be applied here due to the inability of a user to introduce the par-
and coe�cient calculation. The terms )8 = 9 58 9 are paired with
titioning of the processed domain, form �nite-di�erence schema
real-value coe�cients obtained from the optimization subproblem
without a priori knowledge of an equation, proposed by evolution-
of Eq. 3. Finally, the equation coe�cients are detected by linear
ary algorithm. An automatic solving method for candidate equation
regression.
(viewed as in Eq. 6) quality evaluation is introduced in [9] to work
’ ÷ ÷ around this issue.
U 0 = arg min (|| U80 58 9 5< 9 || 2 + _||U 0 || 1 ) (3)
U
8, 8<< 9 9
&B>; ( 0 (D)) = ||D D ||= ! min (5)
In the initialization of the algorithm equation graphs are ran- U 8 C8 9
domly constructed for each individual from the sets of user-de�ned ’ ÷
tokens with a number of assumptions about the structures of the 0
(D) = 0 : 0 (D) = U8 58 9 = 0 (6)
“plausible equations”. 8 9
235
Comparison of Single- and Multi- Objective Optimization �ality for Evolutionary Equation Discovery GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal
While both quality metrics Eq. 4 and Eq. 5 in ideal conditions consumption.10 independent runs are conducted with each setup.
provide decent convergence of the algorithm, in the case of the The main equation quality indicator in our study is the statistical
noisy data, the errors in derivative estimations can make di�erential analysis of the objective function mean (` = ` (& ( 0 ))) and variance
operator discrepancy from the identity (as in problem in Eq. 4) an f 2 = (f (& ( 0 ))) 2 among the di�erent launches.
unreliable metric. Applying the automatic solving algorithm has The �rst equation was the wave equation as on Eq. 8 with the
high computational cost due to training a neural network to satisfy necessary boundary and initial conditions. The equation is solved
the discretized equation and boundary operators. with the Wolfram Mathematica software in the domain of (G, C) 2
As the single-objective optimization method for the study, we [0, 1] [0, 1] on a grid of 101 101. Here, we have employed
> >
have employed a simple evolutionary algorithm with a strategy that numerical di�erentiation procedures.
minimizes one of the aforementioned quality objective functions.
Due to the purposes of experiments on synthetic noiseless data, the m 2D m 2D
= 0.04 2 (8)
discrepancy-based approach has been adopted. mC 2 mG
The algorithm’s convergence due to the relatively simple struc-
2.4 Multi-objective optimization application ture was ensured in the case of both algorithms: the algorithm
As we stated earlier, in addition to process representation, the proposes the correct structure during the initialization or in the
conciseness is also a valuable for regulating the interpretability initial epochs of the optimization. However, such a trivial case can
of the model. Thus the metric of this property can be naturally be a decent indicator of the “ideal” algorithm behaviour. The values
introduced as Eq. 7, with an adjustment of counting not the total of examined metrics for this experiment and for the next ones are
number of active terms but the total number of tokens (:8 for 8 C⌘ presented on Tab. 1.
term).
Table 1: Results of the equation discovery
’
⇠( 0
(D)) = #( ) =
0
:8 ⇤ 1U8 <0 (7)
8 metric method wave Burgers KdV
In addition to evaluating the quality of the proposed solution ` single-objective 5.72 2246.38 0.162
from the point of the equation simplicity, multi-objective enables multi-objective 2.03 1.515 16.128
the detection of systems of di�erential equations, optimizing quali- f2 single-objective 18.57 4.41 ⇤ 107 8.9 ⇤ 10 3
ties of modeling of each variable. multi-objective 0 20.66 ⇡ 10 13
While there are many evolutionary multi-objective optimiza-
tion algorithms, MOEADD (Multi-objective evolutionary algorithm
based on dominance and decomposition) [6] algorithm has proven The statistical analysis of the algorithm performance on each
to be an e�ective tool in applications of data-driven di�erential equation is provided in Fig. 1.
equations construction. We employ baseline version of the MOEADD Another examination was performed on the solution of Burgers’
from the aforementioned paper with the following parameters: PBI equation, which has a more complex, non-linear structure. The
penalty factor \ = 1.0, probability of parent selection inside the problem was set as in Eq. 9, for a case of a process without viscosity,
2
sector neighbourhood X = 0.9 (4 nearest sector are considered as thus omitting term a mmCD2 . As in the previous example, the equation
“neighbouring”) with 40% of individuals selected as parents. Evo- was solved with the Wolfram Mathematica toolkit.
lutionary operator parameters are: crossover rate (probability of
a�ecting individual terms): 0.3 and mutation rate of 0.6.The result mD mD
+D =0 (9)
of the algorithm is the set of equations, ranging from the most sim- mC mG
m= D = 0) to the highly
plistic constructions (typically in forms of mG Derivatives used during the equation search were computed
analytically due to the function not being constant only on small
=
:
complex equations, where extra terms probably represents the noise domain.
components of the dynamics. The presence of other structures that have relatively low opti-
mized function values, such as DG0 DC0 = DCC
00 , makes this case of data
3 EXPERIMENTAL STUDY rather informative. Thus, the algorithm has a local optimum that is
This section of the paper is dedicated to studying equation dis- far from the correct structure from the point of error metric.
covery framework properties. As the main object of interest, we The �nal set-up for an experiment was de�ned with a non-
designate the di�erence of derived equations between single- and homogeneous Korteweg-de Vries equation, presented in Eq. 10.
multi-objective optimization launches. The validation was held The presence of external tokens in separate terms in the equation
on the synthetic datasets, where modelled dependent variable is makes the search more di�cult.
obtained from solving an already known and studied equation.
The tests were held on three cases: wave, Burgers and Korteweg- mD mD m 3D
+ 6D + = cos C sin C (10)
de Vries equations due to unique properties of each equation. The mC mG mG 3
algorithms were tested in the following pattern: 64 evolutionary The experiment results indicate that the algorithm may detect
iterations for the single-objective optimization algorithm and 8 the same equation in multiple forms. Each term of the equation
iterations of multi-objective optimization for the populations of 8 may be chosen as the “right-hand side” one, and the numerical error
candidate equations, which resulted in roughly similar resource with di�erent coe�cient sets can also vary.
236
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal M. Maslyaev, and A. Hvatov
104
102
101
100 10 1
2
10
6 × 100
4
10
0 6
4 × 10 10
8
3 × 10 0 10
10
10
2
2 × 100 10
Single Objective Multi-Objective Single Objective Multi-Objective Single Objective Multi-Objective
Figure 1: Resulting quality objective function value, introduced as Eq. 6, for single- and multi-objective approaches for (a) wave
equation, (b) Burgers equation, and (c) Korteweg-de Vries equation
4 CONCLUSION limit, with active learning and control. Proceedings of the Royal Society A 478,
2260 (2022), 20210904.
This paper examines the prospects of using multi-objective opti- [3] Han Gao, Matthew J Zahr, and Jian-Xun Wang. 2022. Physics-informed graph
mization for the data-driven discovery of partial di�erential equa- neural galerkin networks: A uni�ed framework for solving pde-governed forward
and inverse problems. Computer Methods in Applied Mechanics and Engineering
tions. While initially introduced for handling problems of deriving 390 (2022), 114502.
systems of partial di�erential equations, the multi-objective view [4] L Gao, Urban Fasel, Steven L Brunton, and J Nathan Kutz. 2023. Convergence of
of the problem improves the overall quality of the algorithm. The uncertainty estimates in Ensemble and Bayesian sparse model discovery. arXiv
preprint arXiv:2301.12649 (2023).
improved convergence, provided by higher candidate individual [5] Hisao Ishibuchi, Yusuke Nojima, and Tsutomu Doi. 2006. Comparison between
diversity, makes the process more reliable in cases of equations single-objective and multi-objective genetic algorithms: Performance comparison
with complex structures, as was shown in the examples of Burgers’ and performance measures. In 2006 IEEE International Conference on Evolutionary
Computation. IEEE, 1143–1150.
and Korteweg-de Vries equations. [6] Q. Zhang K. Li, K. Deb and S. Kwong. 2015. An Evolutionary Many-Objective
The previous studies have indicated the algorithm’s reliability, Optimization Algorithm Based on Dominance and Decomposition. in IEEE
Transactions on Evolutionary Computation) 19, 5 (2015), 694–716. https://doi.org/
converging to the correct equation, while this research has proposed 10.1109/TEVC.2014.2373386
a method of improving the rate at which the correct structures are [7] Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. 2021. DeepXDE:
identi�ed. This property is valuable for real-world applications A deep learning library for solving di�erential equations. SIAM Rev. 63, 1 (2021),
208–228.
because incorporating large and complete datasets improves the [8] Mikhail Maslyaev and Alexander Hvatov. 2021. Multi-Objective Discovery of PDE
noise resistance of the approach. Systems Using Evolutionary Approach. In 2021 IEEE Congress on Evolutionary
The further development of the proposed method involves intro- Computation (CEC). 596–603. https://doi.org/10.1109/CEC45853.2021.9504712
[9] Mikhail Maslyaev and Alexander Hvatov. 2022. Solver-Based Fitness Function
ducing techniques for incorporating expert knowledge into the for the Data-Driven Evolutionary Discovery of Partial Di�erential Equations. In
search process. This concept can help generate preferable can- 2022 IEEE Congress on Evolutionary Computation (CEC). IEEE, 1–8.
[10] Mikhail Maslyaev, Alexander Hvatov, and Anna V Kalyuzhnaya. 2021. Partial
didates or exclude infeasible ones even before costly coe�cient di�erential equations discovery with EPDE framework: application for real and
calculation and �tness evaluation procedures. synthetic data. Journal of Computational Science (2021), 101345.
[11] Daniel A Messenger and David M Bortz. 2021. Weak SINDy for partial di�erential
equations. J. Comput. Phys. 443 (2021), 110525.
5 CODE AND DATA AVAILABILITY [12] Lizhen Nie and Veronika Ročková. 2022. Bayesian Bootstrap Spike-and-Slab
The numerical solution data and the Python scripts, that reproduce LASSO. J. Amer. Statist. Assoc. 0, 0 (2022), 1–16. https://doi.org/10.1080/01621459.
2022.2025815
the experiments, are available at the GitHub repository 1 . [13] M Raissi, P Perdikaris, and GE Karniadakis. 2017. Physics informed deep learning
(Part II): Data-driven discovery of nonlinear partial di�erential equations. arXiv
ACKNOWLEDGEMENTS preprint arXiv:1711.10566 (2017). https://arxiv.org/abs/1711.10566
[14] Mikhail Sarafanov, Valerii Pokrovskii, and Nikolay O Nikitin. 2022. Evolutionary
This research is �nancially supported by the Ministry of Science Automated Machine Learning for Multi-Scale Decomposition and Forecasting of
and Higher Education, agreement FSER-2021-0012. Sensor Time Series. In 2022 IEEE Congress on Evolutionary Computation (CEC).
IEEE, 01–08.
[15] Hayden Schae�er. 2017. Learning partial di�erential equations via data discovery
REFERENCES and sparse optimization. Proc. R. Soc. A 473, 2197 (2017), 20160446.
[1] H. Cao, L. Kang, Y. Chen, et al. 2000. Evolutionary Modeling of Systems of [16] H. Schae�er, R. Ca�isch, C. D. Hauck, and S. Osher. 2017. Learning partial
Ordinary Di�erential Equations with Genetic Programming. Genetic Program- di�erential equations via data discovery and sparse optimization. Proceedings
ming and Evolvable Machines 1 (2000), 309–337. https://doi.org/doi:10.1023/A: of the Royal Society A: Mathematical, Physical and Engineering Science (2017).
1010013106294 https://doi.org/473(2197):20160446
[2] Urban Fasel, J Nathan Kutz, Bingni W Brunton, and Steven L Brunton. 2022. [17] Pongpisit Thanasutives, Takashi Morita, Masayuki Numao, and Ken ichi Fukui.
Ensemble-SINDy: Robust sparse model discovery in the low-data, high-noise 2023. Noise-aware physics-informed machine learning for robust PDE discovery.
Machine Learning: Science and Technology 4, 1 (feb 2023), 015009. https://doi.org/
1 https://github.com/ITMO-NSS-team/EPDE_GECCO_experiments 10.1088/2632-2153/acb1f0